Lead story / windows
Windows Zero-Day Lets Low-Privilege Accounts Hijack Admins
A researcher dropped working exploit code the same day Microsoft pushed its largest-ever Patch Tuesday, targeting the Windows User Profile Service.
Tagged / security
security.
The Revision stories tagged security — 362 stories on the wire, decoded into one clear voice.
All stories tagged security
Security
Suno Breach Exposed Code Detailing How It Scraped Songs
A November hack of AI music startup Suno revealed source code that reportedly shows the company scraped millions of songs without permission.
Security
292 Fake GitHub Repos Delivered a Broad-Reach Infostealer
ArcticWolf found hundreds of spoofed repositories pushing a BoryptGrab variant that hits browsers, crypto wallets, and Chrome's App-Bound Encryption.
Security
Old Bootloader Flaws Let Attackers Bypass UEFI Secure Boot
ESET found 11 Microsoft-signed shim bootloaders with ancient flaws that let attackers disable Secure Boot on potentially billions of x86 PCs.
Security
Microsoft Patches a Record 622 CVEs in July Update
AI-assisted bug hunting helped Microsoft triple its previous Patch Tuesday record, though a Dell incompatibility is already causing headaches.
Security
US Sanctions First VPN Admin for Ransomware Aid
Treasury's OFAC designated Ukrainian operator Dmytro Rashevskyi and a Belarusian cryptor seller, targeting the supply chain behind ransomware attacks.
Security
Researcher Tricks Claude Into Exposing Stored User Data
A security researcher demonstrated that Claude's memory feature can be manipulated into revealing private information users stored in the AI.
Security
Tailscale SSH Argument Bug Let Attackers Gain Root Access
Tailscale's bulletin TS-2026-009 describes an argument-handling flaw in its SSH feature that could be used to reach root on connected systems.
Security
Secure Boot Has Had a Gaping Hole for 13 of Its 14 Years
ESET researchers found 11 signed shim images Microsoft never revoked, letting anyone bypass Secure Boot on Windows and Linux machines.
Security
Grok Build Was Sending Full Git Repos to xAI's Cloud
A security researcher found the coding CLI uploaded entire repositories, including secrets and API keys, at 27,800 times the data a task actually required.
Security
Microsoft Patches Record 570 Windows Flaws, Two Exploited Now
July Patch Tuesday triples the previous record and includes two actively exploited privilege-escalation zero-days and a publicly disclosed BitLocker bypass.
Security
ModHeader Pulled from Chrome and Edge After Hidden Spyware Found
A hidden spyware SDK in ModHeader v7.0.18 silently collected visited domains and routed them to a Chinese-owned server for 1.6 million users.
Security
Telegram Links Went Dark After a VPN Sanctions Snafu
A US Treasury crackdown on a ransomware-linked VPN caught Telegram's t.me link shortener in the crossfire when a domain provider cut access too broadly.
Security
Microsoft Upgrades Defender After ShinyHunters OAuth Campaign
Microsoft is adding richer telemetry and OAuth governance to Defender for Cloud Apps after a year-long Salesforce attack hit up to 700 organizations.
Security
Cloudflare Now Flags When DNSSEC Validation Is Quietly Bypassed
After Albania's top-level domain broke DNSSEC and went dark for hours, Cloudflare added a new error code so clients can tell when security checks were skipped.
Security
CISA Flags Active SharePoint Exploits, Urges Patching
Three SharePoint Server vulnerabilities are being actively exploited for remote code execution and credential theft, with two more flagged as high-risk.
Security
Researchers Trick AI Chatbots by Faking Their Own Chat History
A new jailbreak technique exploits AI models' blind trust in their own past messages, bypassing safety filters by forging the conversation history sent via API.
Security
AI Agents Used to Red-Team Other AI Agents
Researchers built a system that autonomously discovers and maps reusable vulnerabilities in production LLM agents like Claude Code and Codex.
AI
Poisoned Data Can Turn AI Research Agents Into Fraud Machines
A new study finds that corrupting public datasets fools frontier AI research agents nearly half the time, with a detection rate of just 6%.
AI
Simple Prompt Tricks Break Google's MedGemma Safety Rules
A new benchmark finds that rephrasing medical questions as board exam items or doctor queries bypasses MedGemma-4B's guardrails 38% of the time.
AI
Persistent Pressure Breaks Every CLI Agent Tested, Study Finds
A new auditing framework stress-tested frontier CLI agents on illegal tasks and found compliance reaching 100% under sustained adversarial pressure.
AI
Least Privilege Is Not Enough for AI Agents, Researchers Argue
A new paper argues that least privilege fails for AI agents that can combine and amplify permissions, and proposes a formal replacement.
Security
Russia's FSB Is Using Your Router as Cover
A joint advisory from CISA and allies warns that Russian state hackers keep hijacking home and small office routers to mask attacks on critical infrastructure.
Security
Apple Says Ex-Employee Grabbed Files After Jumping to OpenAI
A former Apple staffer allegedly used a post-departure bug to pull confidential files from Apple's network before landing at OpenAI.
Security
Ransomware Negotiator Who Played Both Sides Gets 70 Months
Angelo Martino posed as a victim advocate while funneling ransomware payments back to attackers, costing victims more than $75 million.
Security
Ransomware Negotiator Who Worked for the Attackers Gets 70 Months
Angelo Martino, hired to help victims survive ransomware attacks, was secretly feeding BlackCat intelligence and helping infect clients.
Security
A single edit permission was enough to hijack Google Dialogflow agents
Varonis found that one chatbot edit right let attackers plant code, steal credentials, and silently compromise every agent in a Cloud project.
Security
AI-written malware evades antivirus by being one of a kind
A Huntress investigation found a PowerShell script almost certainly generated by an AI chatbot — and its uniqueness is exactly what makes it dangerous.
Security
Defenders Use Prompt Injection to Shut Down AI Attackers
Tracebit researchers found that planting prompt injections next to AWS secrets can halt AI hacking agents by triggering their own safety guardrails.
Security
Russia's FSB Is Raiding Routers. Yours Might Be Next.
A 20-agency joint advisory warns that FSB hackers are systematically stealing router configs via default SNMP credentials across critical infrastructure.
Security
Lidl Data Breach Hits Online Shoppers Across Europe
A third-party IT provider was breached, exposing names, emails, and dates of birth for Lidl customers in at least three countries.
AI
SafetyNet Cracks Open How AI Models Dodge Their Own Watchdogs
New research exposes the mechanics behind white-box monitor evasion and proposes an ensemble defense that hits near-perfect detection scores.
AI
China Flags Claude Code as Security Risk, Boosting Local Rivals
Beijing's cybersecurity alert targeting Anthropic's coding tool is accelerating a migration already under way toward domestic Chinese alternatives.
Security
Startup Sues Koi Security Over AI Hallucination in Threat Report
MeetingTV claims Koi's Wings platform invented a fake browser extension to tie the video startup to Chinese espionage, then published without verification.
Security
A PNG Image Tricked AI Coding Tools Into Leaking Data
New research shows a prompt injection hidden inside a PNG can manipulate AI coding assistants into exposing sensitive information.
Security
CISA Had No Incident Playbook When It Got Hacked
The agency that tells organizations how to respond to breaches was writing its own response plan in real time during a May security incident.
Security
Nebusec Claims 15-Year Linux Stack Bug Across All Distros
A security research firm says a stack use-after-free bug called GhostLock hides in every Linux distribution, but its age claim doesn't hold up.
Security
Ransomware Negotiator Took Kickbacks While Cutting Deals for Victims
Angelo Martino, a negotiator hired to lower ransom payments, fed confidential client data to BlackCat affiliates and pocketed a cut of the inflated ransoms.
Security
AssuranceAmerica Breach Hits 6.9 Million Driver Records
A social engineering attack on an employee gave criminals access to driver's license numbers and other personal data from the auto insurer.
Security
Phishers Used Meta's Own Email System to Steal Business Credentials
Attackers exploited Meta's business-to-business email infrastructure to send convincing phishing emails that bypassed spam filters entirely.
Security
Fake Google Notes Extension Hijacks Crypto Clipboard
McAfee identified "Silent Swap," a Chromium extension masquerading as Google Notes that swaps crypto wallet addresses mid-paste to redirect funds.
AI
AI Agents That Remember You Have a Forgetting Problem
New research finds that deleting data from persistent AI agent memory often leaves recoverable traces, even after a user asks for erasure.
AI
Hidden Traps That Make AI Agents Forget Your Content
A new framework called CAPE exploits the compression step inside AI agents to cause information loss, without changing anything a human reader can see.
Security
Prismata Aims to Stop Prompt Injection in Web Agents
A new defense system called Prismata limits what AI web agents can see and do, reducing prompt injection attacks without requiring any developer setup.
Security
CodeTracer Hunts the Source of Backdoored AI Code Completions
A new forensic framework traces malicious code suggestions back to the poisoned training data that planted them.
Security
Adversarial Decoys Let Patch Attacks Slip Past ViT Defenses
Researchers found that a second, decoy patch can redirect the attention-based filters that protect Vision Transformers, leaving the real attack undetected.
AI
Inside the Jailbreak: How LLMs Get Rerouted
New research maps the internal computation paths that adversarial prompts hijack, offering a more surgical view of why AI safety guardrails fail.
AI
Overthinking Trick Forces AI Models to Leak Training Secrets
Researchers found that amplifying a model's reasoning weights can surface hidden information up to 10 times more often than standard auditing methods.
Security
Windows Defender Patch Trades One Bug for a Disk-Filling Risk
Microsoft's fix for a remote-code-execution zero-day in its Malware Protection Engine may cause Windows machines to consume all available disk space.
Security
Tenda Routers Carry a Hardcoded Backdoor with No Fix in Sight
A critical 9.8-severity flaw lets attackers log in to multiple Tenda router families with a hidden password, and the vendor has not responded.
Security
Accenture Confirms 35GB Breach from Azure DevOps Repos
A hacker known as 888 claims to have stolen source code, RSA keys, and Azure credentials from Accenture's DevOps infrastructure.
Security
Suspected Chinese Spies Hit University Mail Servers
A group tracked as UNK_MassTraction has been raiding Roundcube mail servers at US and Canadian universities since May, stealing credentials from research staff.
Security
Researchers tricked GitHub Copilot by hiding harm in a workflow
Alan Turing Institute researchers found that spreading a harmful request across a coding workflow bypasses Copilot's direct-chat safety filters.
Security
Nextcloud Exposed 367K Records in Misconfigured Database
An unprotected ElasticSearch cluster left Nextcloud employee emails, client contracts, and internal scripts visible to anyone who found it.
Security
Hard-coded Credentials Found in Schneider Electric Grid Relays
A CVSS 5.3 flaw in Schneider Electric's MiCOM Px40 protection relays lets unauthenticated attackers pull device data via SNMP using baked-in credentials.
Security
Critical Flaw in OpenPLC v3 Lets Attackers Run Code on ICS
A CVSS 9.9 file-write bug in OpenPLC v3's web UI can be escalated into full native code execution — and the only fix is upgrading to v4.
Security
Your Spotify Playlists Can Leak Your Age, Gender, and Habits
Researchers built a tool that infers personal attributes from public playlists using deep learning — and proposes a defense that only partly works.
Security
AnyPoC Turns LLM Bug Reports Into Executable Proof-of-Concept Tests
A multi-agent framework validated 108 developer-confirmed bugs across 12 major codebases, with 46 of its test cases adopted as official regression tests.
Security
New Jailbreak Method Hits 96.8% Success Against Safety-Aligned LLMs
A gradient-based attack called NTA drops the fixed-target requirement and nearly doubles the success rate of leading jailbreak techniques.
AI
A Better Way to Grade AI Agent Hacks
Researchers say the standard pass-fail metric for agent red-teaming hides how much damage a successful attack actually causes.
Security
A New Attack Exploits the Math Inside Vision-Language Models
Researchers found that targeting the spectral structure of a model's internal layers makes adversarial attacks more effective than existing methods.
AI
Multi-Agent AI Attacks Slip Past Per-Agent Monitors
A new study finds that coordinating AI agents across a shared codebase defeats the per-instance oversight most deployments rely on.
AI
New Framework Detects Rogue AI Agents Without Graph Maps
AcMAS reads agents' internal activation states to catch stealthy attacks, gaining +0.22 F1 in sync and +0.55 F1 in async settings over graph-based tools.
Security
AI Agents With Long-Term Memory Can Be Poisoned
A new attack called GhostWriter can implant false memories in AI personal agents at near-universal injection rates, with no guard in place to stop it.
AI
A Math Trick Makes Object Detectors Harder to Fool
Researchers built Lipschitz constraints directly into a detection architecture, cutting adversarial vulnerability without leaning on attack-specific training.
Security
JADEPUFFER Is the First LLM-Run Ransomware Attack
Sysdig researchers say an AI agent autonomously exploited a Langflow flaw, encrypted data, and issued ransom demands — with no human operator involved.
Security
China's NVDB Alleges Claude Code Sent User Data Without Consent
Beijing's NVDB alleges Claude Code versions 2.1.91-2.1.196 secretly harvested location and identity data — a claim Anthropic has not yet addressed.
Security
GitHub's AI Agent Can Be Talked Into Leaking Private Repos
Security firm Noma Labs found that a politely worded issue ticket is enough to trick GitHub's AI coding agent into exposing private repositories.
Security
GitHub's AI Agent Was Tricked into Leaking Private Repos
Security researchers found they could manipulate GitHub's AI agent into exposing private repository data through prompt injection.
AI
A Gateway That Stops Deceived AI Agents From Acting Out
aiAuthZ moves authorization off the agent entirely, cutting residual attack success to zero across 15 models with under 0.03 ms of added latency.
AI
AI Lie Detection Gets a Multicultural Dataset and Audit Trail
Researchers released T4-Deception, the largest real-world deception detection dataset, plus a model that explains its reasoning across cultures.
AI
Deepfake Audio Detectors Can't Handle Cleaned-Up Real Voices
A new study finds that speech restoration and voice quality tools break the binary logic that most audio anti-spoofing systems rely on.
Security
Researchers Built a Benchmark to Test If LLMs Can Escape Sandboxes
A new open benchmark finds that frontier LLMs can exploit container vulnerabilities when given shell access — raising questions about AI agent deployments.
AI
StepShield Finds AI Safety Monitors Alert Too Late to Matter
A new benchmark reveals that pattern-based AI agent guardrails catch bad behavior but fire alerts so late they resemble random timing.
AI
SecureCode Targets the Vulnerable Code AI Assistants Keep Writing
A new open dataset pairs web and AI/ML security examples to help train coding models that stop producing vulnerable output.
Security
Thermal-Airflow Trick Fools Infrared AI at a 48% Rate
Researchers found that simulated airflow turbulence can reliably mislead vision-language models analyzing infrared satellite imagery.
AI
Fine-Tuning AI Decompilers Can Make Them Worse
A new benchmark study finds that fine-tuning large models for Dart binary decompilation often hurts accuracy, even when simpler metrics look fine.
AI
RoME Makes AI Models Harder to Fool Across Attack Types
A new training method from arXiv paper 2607.06109 uses a mixture of low-rank experts to close the gap between defenses against different adversarial threats.
AI
A New Model Hunts Influence Ops Without Training Labels
Researchers built an unsupervised detector that flags coordinated inauthentic accounts using timing patterns and LLM-scored post content.
AI
Removing Safety Filters From LLMs Boosts Vuln Analysis - at a Cost
A new study compares aligned and refusal-stripped models from the same lineage and finds abliterated versions perform meaningfully better on security tasks.
Security
MCP Tool Approval Can Be Spoofed Eight Different Ways
Researchers found eight techniques to smuggle malicious instructions into AI coding agents via MCP, with one making the payload invisible to human reviewers.
AI
Vision Datasets Carry Hidden Attack Surfaces
Researchers find that ordinary image datasets like ImageNet contain statistical patterns that can skew model predictions without any deliberate tampering.
AI
A Smarter Way to Keep Security Models Stable After Retraining
Researchers propose a calibration method that keeps false-positive rates consistent across model updates, sparing downstream teams from constant recalibration.
AI
A Forensics-Ready AI System That Stays On-Premise
Researchers built Charlie, a multi-agent RAG system designed to handle digital forensic evidence without sending data to the cloud.
AI
A Paper Proposes Cryptographic Receipts for AI Agent Actions
Researchers want AI agents to prove every step they took was authorized, logged, and replayable — not just that their final output looks right.
AI
A New Tool Revokes Coding Agent Permissions When the Job Is Done
Researchers propose PORTICO, a reference monitor that strips coding agents of tool access as soon as each subgoal closes, not at session end.
AI
Physics Formulas Help AI Explain IoT Anomalies
A new framework borrows from statistical mechanics to attribute fault and abnormal behavior in industrial IoT systems without needing a full causal map.
AI
One Attack to Test All AI Models for Training Data Leaks
Researchers built a single membership inference framework that works across text, image, and multimodal AI models without needing access to model internals.
Security
A Backdoor That Hides in Plain Math
Researchers have built a supply-chain attack that plants an undetectable backdoor in pre-trained image classifiers, with a formal proof to back the claim.
AI
Vision Transformers Beat CNNs at Catching Deepfakes
An ensemble of fine-tuned vision transformers hit 96.77% AUC on a tough deepfake benchmark, outpacing the previous best detector by over seven points.
AI
Kernel-Level Safety for AI Tool Calls
A researcher built OS-level governance for MCP tool calls, arguing that app-layer guardrails are too easy to bypass.
AI
Symbolic Guardrails Beat Guesswork for AI Agent Safety
A new study argues that rule-based checks, not more neural training, can enforce 74% of AI agent security requirements with minimal overhead.
Security
Tiny Local Model Matches Frontier AI on Linux Privilege Escalation
Researchers trained a 4B-parameter model to crack Linux privilege escalation at 93% accuracy, cutting inference costs 80x versus cloud-based frontier AI.
AI
A Context Filter Claims to Cut LLM Jailbreaks by 92%
Researchers propose a plug-and-play input filter that strips adversarial context from prompts before they reach a model, without retraining it.
Security
A Steering Image Can Quietly Rewire an AI's Behavior
Researchers found that a single optimized image can shift a vision-language model's behavior by 25%, no text prompts or model internals needed.
AI
kAgent Automatically Patches Linux Kernel Fuzz Bugs
Researchers built kAgent, an AI workflow agent that patches Linux kernel crashes from fuzzer output, resolving up to 65% of bugs with file location hints.
Security
Audio Deepfake Detectors Crumble Under Compression
A new study stress-tests 10 detection models against 18 real-world corruptions and finds most fall apart when audio is compressed or modified.
Security
Inside LLM Refusals: A Faster Path to Jailbreaks
Researchers used mechanistic interpretability to map how LLMs decide to refuse prompts - then exploited that map to bypass safety guardrails in seconds.
AI
Verifying AI Robustness Without Exposing the Model
SecureCROWN lets two parties jointly check neural network robustness guarantees without either side revealing its private data.
Security
AI Agents Have a New Attack Vector Researchers Are Calling ADI
A new class of injection attack tricks AI agents into treating malicious data as trusted, bypassing defenses built for older prompt injection methods.
Security
Attackers Can Forge an AI Agent's Memory of Its Own Reasoning
A new attack called FARMA poisons the reasoning history of LLM agents, bypassing existing defenses — and researchers say a layered pipeline can stop it cold.
Security
New Rust Security Benchmark Catches LLMs Struggling to Localize Bugs
RustMizan tests AI agents on real, compilable Rust code and finds that pinpointing a vulnerable line is still far beyond their reach.
Security
A Few Bad Training Samples Can Silently Break a Robot
Researchers found that just three poisoned episodes out of 320 can reduce a robot's task success rate to zero — with no visible signs during normal use.
AI
Text-to-SQL Models Break Easier Than Anyone Admitted
A new automated framework called SAGE finds that LLMs converting plain English to database queries fail in ways static testing never caught.
Security
DualView Blocks Stored Prompt Injection in Local AI Agents
Researchers present a two-view architecture that tracks malicious data across file systems and networks, fixing a gap that prior defenses left open.
Security
New Benchmark Finds Reworded Poison Docs Beat RAG Defenses
Researchers show attackers can boost RAG hijacking fivefold by rewording poisoned documents, while standard success-rate metrics miss most of the damage.
AI
AI That Writes Access Rules and Then Checks Its Own Work
AutoCedar uses a verifier loop to convert natural-language requirements into Cedar policies that are formally checked before anyone ships them.
AI
A Lightweight Tool to Catch What Fine-Tuning Hides
SAR, a new LoRA adapter, lets practitioners ask a fine-tuned model what hidden behaviors it actually learned — and get a reliable answer.
AI
New Model Spots Faces Behind Masks More Accurately
A transformer-based system uses eye and brow geometry to match masked faces, hitting 97% accuracy where older CNN methods stall out.
AI
A New System Blocks Unsafe AI Agent Tool Chains Before They Run
Researchers propose a two-phase security framework that catches policy violations created by combining AI tools, not just by individual ones.
AI
Resemble AI's Deepfake Detector Passes a Fairness Audit
A large-scale study finds DETECT-3B-Omni's accuracy varies by at most 2 points across speaker age, gender, region, and content type.
Security
AI System Spots Malicious Python Packages via Code Graphs
A new framework pairs large language models with graph neural networks to flag malicious PyPI packages without requiring human review.
AI
AI for OSINT Has a Hallucination Problem Nobody Is Measuring
A survey of 74 studies finds that AI-assisted intelligence tools rarely validate their own outputs, leaving a critical reliability gap in cyber investigations.
AI
AI Surveillance System Cuts False Alarms from 52% to 4%
A new multi-task deep learning framework handles face recognition, weapon detection, and fire alerts on a single GPU while slashing false-alarm rates.
Security
New AI Model Targets Intrusion Detection in IoT Networks
Researchers combined graph neural networks with Kolmogorov-Arnold Networks to catch low-frequency attacks that standard detectors miss.
AI
A Smarter Way to Hide Your Private Data from AI Chatbots
PromptPET picks the right obfuscation technique per prompt to keep your data private without making chatbot responses useless.
AI
AI Explainability Tools Can Leak Training Data - Here is a Fix
Researchers propose a training-time defense that closes a privacy gap opened by the very tools designed to make AI more transparent.
Security
For AI Security Agents, the Client Beats the Model
A new study using HexStrike-AI as a testbed found that the client driving an LLM matters more than the model itself in security tool orchestration.
Security
JavaVulBench Tests 12 AI Models on Real Java Security Flaws
A new open benchmark pits CodeBERT against GPT-4o and Claude Sonnet 4 on 30,600 Java methods tied to 1,740 CVEs - with a built-in cheat-detection mode.
Security
Hijacking 10% of Vision Tokens Can Break AI Model Accuracy by 88%
New research shows that intercepting a small slice of data in transit between edge devices and cloud AI servers can nearly destroy a model's accuracy.
Security
How Domain-Specific Abliteration Targets AI Safety Refusals
Researchers tested abliteration on 24 open-source LLMs and found they could strip cybersecurity-domain refusals without removing broader safety behavior.
AI
Microsoft's Fix for Unreliable AI-Generated Security Artifacts
A framework deployed in Microsoft Sentinel flips the LLM reliability problem: stop trying to make AI generate perfectly, and validate obsessively instead.
AI
How Poisoned Web Pages Can Hijack AI Research Agents
A new attack method called FORGE shows how adversarial documents can silently corrupt the reports produced by deep research AI agents.
AI
A Cryptographic Lock to Keep AI Agents in Their Lane
Researchers show that agent authority can be guaranteed by architecture, not just training — using cryptographic identity to enforce hard limits.
Security
AI Ransomware Detector Knows When to Ask a Human
A new multi-agent framework called SABRE uses uncertainty scoring to decide when to block ransomware automatically and when to escalate to an analyst.
AI
Biology-Inspired Patterns to Tame Rogue AI Agents
Researchers map five biological control motifs to agentic AI design patterns, offering a typed framework to curb hallucinations and prompt injection.
AI
A New Method Reads LLM Reasoning Without Labels
Researchers built an unsupervised framework that extracts how language models reason internally, no human-labeled training data required.
Security
An AI Agent Ran a Full Ransomware Attack Solo
An autonomous AI agent reportedly completed a full ransomware attack on its own, though researchers, targets, and methods remain unnamed.
Security
AI Agent Ran a Full Ransomware Attack With No Human Help
Security firm Sysdig says it documented the first end-to-end ransomware attack executed entirely by an AI agent, with no human operator involved.
AI
Web AI Agents Fall for Prompt Injection 25% of the Time
A new benchmark finds that hidden instructions in web interfaces can hijack autonomous AI agents, with success rates varying widely across frontier models.
Security
ContraFix Hits 92% on Security Bug Repairs Using Contrastive Analysis
A new agentic framework pairs runtime contrast analysis with a reusable skill base to outpace existing automated vulnerability repair tools by a wide margin.
AI
LLMs Write Cloud Access Policies, But Miss the Mark Half the Time
New research finds reasoning models nail access control policy generation 94% of the time, but standard LLMs get it right less than half the time.
Security
RedCoder Automates Multi-Turn Attacks on Code AI
A new red-teaming agent called RedCoder probes code-generating AI over multi-turn chats to surface vulnerabilities that single-shot tests miss.
AI
A New Method Explains Why Image Models Fail Under Attack
Researchers built I-ASIDE, a tool that uses Shapley values to map which image frequencies make vision models fragile or robust.
AI
SkillFuzz Finds Hidden Agent Risks in Open Skill Marketplaces
A new execution-free testing tool exposes how combining individually safe AI agent skills can produce dangerous unintended behavior at scale.
Security
How Researchers Tell a Guardrail Block from an AI Refusal
A new black-box methodology can detect whether an AI safety guardrail or the model itself triggered a refusal, with near-perfect accuracy.
AI
kNNGuard Skips Fine-Tuning to Screen LLM Prompts
A new guardrail system uses hidden activations and nearest-neighbor search to flag unsafe prompts without any model training.
AI
A New Test Flags Stripped Safety Filters in AI Checkpoints
Researchers built a two-signal audit that identifies abliterated open-weight models before deployment, catching 53 of 57 known cases in a 273-checkpoint test.
Security
New Attack Hides Backdoors in Voice AI Using Timbre
Researchers built a method that embeds hidden triggers in speech models by leaking timbre data at the frame level, slipping past standard defenses.
Security
Adversarial Attacks on Security AI Break More Than Predictions
New research finds that attacks fooling cybersecurity classifiers also destabilize the explanations analysts use to triage alerts.
AI
Researchers Can Guess LLM Architecture Through Restrictive APIs
A new attack called NightVision estimates hidden dimensions, depth, and parameter counts of commercial LLMs using only standard API access.
AI
CPG-PAD Uses Explainable AI to Spot Fake Faces Across Domains
A new framework embeds visual concept guidance into prompt learning to help face recognition systems catch spoofing attacks they have never seen before.
AI
ProvenanceGuard Cuts AI Agent Errors by Tracing Actions to Intent
A new multi-stage pipeline checks whether an AI agent's tool calls trace back to what the user actually asked for, slashing misalignment error rates.
AI
Treating Coding Agents Like Human Engineers Catches More Bugs
A new paper argues that access controls and coding conventions, not complex AI scaffolding, are the cheapest way to keep coding agents in check.
AI
A Guard Rail Framework for AI Running Telecom Networks
Researchers propose a runtime validation layer that checks AI decisions in autonomous telecom networks before they touch live infrastructure.
AI
ElephantAgent Wants to Stop AI Agents From Being Poisoned
A new protocol called ElephantAgent uses cryptographic state tracking to catch when an AI agent's memory or tools have been tampered with.
AI
English-Only Safety Training Leaves LLMs Open to Multilingual Attacks
A new gradient-guided attack exploits gaps in multilingual safety coverage to bypass refusals in leading open-source models at rates above 90%.
Security
Hidden Triggers in Voice AI Can Fool Models and Dodge Defenses
A new attack embeds invisible backdoors in speech classifiers using reinforcement learning, and standard defenses like fine-tuning and pruning don't stop it.
AI
Janus Tests Who Should Really Control Your AI Agent
A new open-source research system shows that no single permission design works for all situations — and that ignoring user fatigue is a serious design flaw.
AI
AI Medical Image Detectors Fooled by Metadata Swaps
Researchers found that vision-language models can flip their fake-image verdicts based on text context alone, a gap with real consequences for clinical AI.
AI
How Web Pages Can Manipulate AI Search Agents
New research shows that coordinated networks of linked pages can steer AI shopping agents toward specific products, bypassing traditional search ranking.
AI
Open-Source Tool Scrubs PII from AI Agent Payments
A new middleware layer intercepts x402 payment requests before transmission, filtering personal data before it reaches servers with no data agreements.
AI
Moire Patterns Could Be the Key to Spotting AI Video
Researchers found that a physics-based optical signature from real cameras is one AI video generators cannot fake.
Security
AI Auditor Knowdit Finds DeFi Bugs Humans Miss
A new agentic system built on historical audit reports caught every high-severity vulnerability in a benchmark of 11 DeFi projects — with zero false positives.
AI
A Cheaper Way to Stop AI Agents From Oversharing
Researchers propose activation-based privacy filters for LLM agents that outperform expensive auxiliary monitor approaches in both speed and protection.
Security
Small Security Models That Punch Above Their Size
CyberPal 2.0, a family of 4B-to-20B-parameter models, beats GPT-4o and o3-mini on threat-investigation tasks while staying a fraction of their size.
Security
Federated Learning Audits Miss a Whole Class of Backdoor Attacks
A new stress-testing framework called FLAT shows that standard accuracy metrics can mask backdoor vulnerabilities that vary by trigger and target.
AI
Microsoft Proposes a Cryptographic Way to Prove LLM Ownership
A new fingerprinting framework ties model outputs to cryptographic proofs, aiming to make AI model theft easier to detect and prosecute.
AI
A New Method Forces Hidden LLM Biases Into the Open
Researchers built a tool that amplifies concealed model biases until they show up in plain text, bypassing the need to know what bias to look for.
AI
AI Agent Skills Have a Supply Chain Problem
A study of 1.43 million AI agent skills found hidden malicious dependencies that single-skill audits miss entirely.
Security
GPT-2 Hidden States Can Leak Your Input Text
A new study shows gradient-based attacks can recover input text from a model's hidden states with up to 97.5% accuracy on short prompts.
Security
IIoT Intrusion Detectors Fail Outside Their Training Networks
A new study finds lightweight ML models for industrial network security collapse when tested on networks they were not trained on.
Security
A New Jailbreak Targets AI Tool-Calling, Not Just Prompts
Researchers found that LLMs using function-calling can be manipulated by faking a moderation audit across multiple turns, bypassing prompt-level safety.
Security
On-device AI Brings New Attack Surfaces to Your Phone
A new research paper maps out how locally deployed AI models on mobile devices open security gaps that cloud-based inference never had.
AI
A New Algorithm Cuts RAG Poisoning Attacks to Near Zero
PRA-RAG uses geometric embedding analysis to reduce retrieval poisoning success rates to 1%, while keeping accuracy at 71% across tested benchmarks.
Security
BioShocking Attack Tricks AI Browsers Into Giving Up Passwords
Security firm LayerX found that framing a credential-theft attack as a game fooled every AI browser it tested into leaking user passwords.
Security
Bad Math Tricks AI Agents Into Stealing Credentials
A 'BioShocking' attack manipulates AI agents by rewarding wrong answers, then redirects them to harvest sensitive login credentials.
Security
Fake Reasoning Tricks AI Into Giving Out Drug Recipes
A new attack called CoT Forgery fools LLMs into treating injected text as their own conclusions, pushing jailbreak success from near zero to 60%.
AI
LLM Safety Guardrails Have a Bypass Built In
Researchers found that LLM safety guardrails sit on a linear 'refusal axis' that can be flipped in about one second, or hardened without retraining.
Security
Semantic Caching in LLMs Has a Collision Problem
A new attack framework shows how the fuzzy matching that makes LLM caches fast also makes them exploitable for response hijacking.
AI
Teaching AI to Learn From Noisy Human Feedback
Researchers have built the first provably robust offline RLHF methods, designed to recover reliable AI behavior even when training data is partially corrupted.
AI
A Smarter Way to Catch AI Models Memorizing Your Data
Researchers built a framework that loops AI outputs back as inputs to expose memorized training data more reliably than existing methods.
AI
AD-CERT Closes the Gap Between AI Robustness and Verifiability
A new training method combines adversarial distillation with formal verification bounds to beat benchmarks without sacrificing accuracy.
AI
The LLM Security Problem Is Bigger Than the Model
A new survey maps vulnerabilities across the full LLM stack — from data collection to autonomous agents — and finds that point defenses are not enough.
AI
One Form to Report AI Flaws to Everyone
FLARE-AI is an open-source reporting tool that lets researchers file a single standardized flaw report and route it to multiple AI developers at once.
Security
A Knowledge Graph That Links CVEs to Attack Behaviors
Researchers built a graph connecting software vulnerabilities to MITRE ATT&CK tactics, giving defenders a structured map from bug to behavior.
Security
New Method Detects Hidden Backdoors in LLMs Before They Fire
Researchers have built a detection framework that finds and reconstructs backdoor triggers in large language models used as classifiers.
Security
New Method Removes LLM Backdoors Without Full Retraining
Researchers used curvature analysis to pinpoint and surgically repair the exact model weights responsible for backdoor behavior in large language models.
Security
LLMs Write Functional Malware With 84% Real-World Overlap
New research tested open-weight language models on PowerShell malware generation and found the output alarmingly close to the real thing.
AI
Blocking Prompt Injections Has a Hidden Cost
A new benchmark shows that defenses against prompt injection attacks often work by suppressing text — breaking tasks like translation in the process.
AI
A Browser Plugin That Catches Audio Deepfakes Locally
Researchers built an on-device audio deepfake detector for journalists who need to verify sources without sending audio to the cloud.
Security
AI agent plugins achieve 100% attack success in new benchmark
A new 406-task benchmark found that malicious plugins in always-on AI agents had a 100% attack success rate, regardless of which language model powered them.
AI
AgentBound Wants AI Agents to Show Their Work
A new runtime framework cryptographically logs every decision an autonomous AI agent makes, giving auditors something to actually verify.
Security
AI Browsers Can Be Tricked Into Ignoring Their Own Rules
New research shows a website can trap an AI browser in a fabricated reality where its safety guardrails stop working entirely.
Security
Apple Moves to Faster Security Patches as AI Cuts Hacker Lead Time
Apple says AI-accelerated attacks forced it to abandon its old scheduled patch cadence in favor of faster, standalone security updates.
Security
AI Models Can Now Hack Servers on Their Own
A new benchmark tests 19 AI models on autonomous penetration attacks, with success rates reaching nearly 70% on realistic targets.
Security
LLMs Fail Code Security Reviews When the Context Changes
A new study finds that framing, anchoring, and author attribution can flip an AI security tool's verdict without changing a single line of code.
Security
AI Agents Auto-Patch Enterprise Networks After Attacks
A new multi-agent framework called COHORT cuts the weeks-long manual process of hardening networks after a cyberattack down to an automated pipeline.
AI
AI Shopping Agents Can Be Gamed by Sellers
A new benchmark shows that sellers who rewrite product listings for AI recommendation engines can boost flawed products into results by up to 83%.
AI
A Math-Based Router That Tunes Out Lying AI Agents
ANTAP routes tasks in multi-agent AI systems using algebraic projections instead of agent self-descriptions, cutting attack success rates to near zero.
AI
ChainCaps Tackles the Hidden Risk in AI Tool Chains
A new MCP proxy called ChainCaps blocks a class of AI agent exploits where each tool step looks safe but the combined effect isn't.
Security
A New Attack Flips AI Safety Filters Across Models
Researchers found that targeting high-entropy decision tokens can bypass safety guardrails in vision-language models and transfer across architectures.
AI
A New Protocol Shields AI Agent Networks from Bad Actors
Researchers propose Self-Anchored Consensus, a decentralized filter that stops rogue AI agents from corrupting multi-agent reasoning networks.
AI
A Simple Trick Exposes What Finetuned AI Models Are Hiding
Researchers found that ranking outputs by perplexity difference between a finetuned model and its base checkpoint reliably surfaces hidden training objectives.
AI
AI Refusal Is a Gate, Not a Wall
Researchers found that safety training in large language models works by routing, not removal — and simple ciphers can collapse it almost entirely.
AI
AI Agents Learn to Solve CAPTCHAs on Their Own
A new model called ReCAP trains AI agents to crack interactive CAPTCHAs by learning from their own mistakes, without losing general web-navigation ability.
AI
Proving AI Guardrails Actually Ran, and the Limits of That Proof
Researchers propose cryptographic proof that AI guardrails ran, but warn that even verifiable safety can be gamed by determined developers.
AI
Fine-Tuning an Edited AI Model Wipes Most of the Edits
New research finds that knowledge edits to large language models decay sharply once fine-tuning begins, cutting both a safety lifeline and a liability.
Security
LLM Rankers Can Be Gamed With Carefully Crafted Text
Researchers built a two-stage attack that slips a target item to the top of AI-generated search rankings using natural-sounding text.
AI
Safety Rules Vanish When AI Agents Run Out of Context
A new benchmark shows LLM agents that obey constraints perfectly can break them at a 59% rate once context compaction quietly drops the rules.
AI
GraphChase Opens Up a Benchmark for AI Cop-and-Robber Games
A new open-source platform lets researchers test algorithms that model how law enforcement intercepts criminals across city road networks.
Security
A Framework for Finding Weak Links in Multi-Agent AI Systems
MESA ranks the communication channels most likely to be exploited in multi-agent AI systems, letting defenders focus scarce resources where it counts.
Security
LLMs Scan Code With the Same Biases Humans Do
A new study finds AI vulnerability detectors can be fooled by author reputation, framing, and prior analysis results — not just by the code itself.
Security
A Smarter Malware Classifier That Handles Obfuscation
Researchers propose a multi-task AI framework that detects packed binaries, classifies malware families, and flags threats with a 97.4% combined detection rate.
AI
Hidden Harmful Training Data Can Slip Past AI Guardrails
Researchers found a way to embed harmful instructions inside seemingly safe fine-tuning samples — and most existing defenses miss it entirely.
Security
A Smarter Way to Catch Network Intruders Without Raw Packets
New entropy-based framework spots intrusions using only flow-level stats, exposing blind spots that standard F1 scores quietly hide.
AI
Fuzzing LLMs Can Shake Loose Hidden Backdoor Behaviors
Researchers found that injecting noise into model weights or activations surfaces concealed unsafe behaviors more reliably than standard sampling techniques.
AI
A New Defense Fills the Gap That Leaves AI Safety Blind
Researchers found that most AI safety defenses miss a whole class of attack, then built a fix that cuts the success rate to zero.
AI
AI Agents Mistake Hearsay for Fact When Memory Gets Compressed
A new study shows memory-compression tools silently upgrade hedged remarks into confident facts that LLM agents act on without question.
AI
A Game-Theory Model for How LLMs Can Deceive and Be Deceived
Researchers built a formal framework to study how language models can be manipulated through carefully chosen prompts, and how to design defenses against it.
Security
MCP Agent Runtimes Have a Security Gap Researchers Want Fixed
A new benchmark finds that connection-layer safeguards alone leave AI agent systems exposed, and proposes a tighter execution-control layer to close the gap.
AI
Researchers Show How to Poison Recommender Fairness
A new reinforcement learning attack method can deliberately worsen bias in recommendation systems, even ones designed with fairness protections built in.
AI
Error-Correcting Codes Could Sharpen AI Decision Accuracy
New research applies repetition codes to hash-based homomorphic AI, claiming arbitrary reductions in validation error for decision tests.
AI
Clustering Trick Cuts Speech Poisoning Attack Rate to Near Zero
Researchers used unsupervised clustering to filter poisoned training data, dropping attack success rates from nearly 100% to almost nothing.
AI
AI Diffusion Model Takes on Wireless Jamming
Researchers propose a framework that uses Brownian bridge diffusion to clean up jammed wireless signals before they reach the decoder.
Security
Privacy Tricks for AI Security Tools Fall Short
A study of 96 fine-tuned small language models found that formal privacy guarantees add little beyond what simpler controls already achieve.
Security
LLM Agents Are Both a Security Risk and a Security Tool
A new survey maps how AI agents create fresh attack surfaces while also powering the offense-defense lifecycle — and argues the two problems inform each other.
Security
AI-Powered Intrusion Detectors Can Be Fooled 92% of the Time
A new packet-level attack method evades neural network intrusion detectors at a 92.78% rate while keeping malicious traffic functionally intact.
AI
AI Agents Can Now Hide Messages That Monitors Cannot Catch
New research shows that agentic coding models can build undetectable steganographic channels using everyday tools like code execution and web search.
AI
AEGIS Framework Spots Adversarial Attacks on Vision AI
A new detection system combining semantic analysis and uncertainty modeling catches six categories of adversarial image attacks with over 90% accuracy.
Security
RL for Bug Hunting in C/C++ Is Still Mostly Theoretical
A systematic review of 21 studies finds reinforcement learning has barely scratched the surface of actual vulnerability detection in C/C++ code.
AI
AI Phone Agents Bought Poison Precursors in Real-Device Tests
A new study found AI phone agents completed harmful tasks at a 68.8% rate, including one that deceived a doctor to obtain a toxic precursor.
AI
A New Adapter Technique Cuts Sneaky AI Backdoors
Inoculation adapters outperform prompt-based methods at suppressing unwanted AI behaviors, though keeping desired traits intact remains unsolved.
Security
New Attacks Show GNNs Leak Sensitive Graph Data
Researchers demonstrate two reconstruction attacks that recover private training data from graph neural networks, even in black-box scenarios.
Security
Poisoned Memories Can Steer AI Agents Wrong
New research shows that corrupting an LLM agent's memory store is enough to flip its answers, even when the questions it receives are perfectly clean.
Security
US Puts $10M Bounty on Russian Group Hacking Signal and WhatsApp
Federal authorities are offering $10 million for leads on a Russian intelligence-linked group that phished thousands of encrypted messaging accounts.
Security
iOS 26.5.2 Fixes 25-Plus Security Flaws
Apple's latest point release patches kernel and WebKit bugs that could expose data — and publishing the details raises the urgency to update now.
Security
KDDI Breach Leaks 14M Credentials Across Six ISPs
A third-party software flaw let hackers into KDDI's systems, exposing email and password data for millions of customers at six Japanese ISPs.
Security
GitHub's Vuln Database Hit a Record in May and Still Fell Behind
The GitHub Advisory Database published 1,560 reviewed advisories in May 2026 — five times its usual output — and still couldn't clear the backlog.
AI
Seven Security Gaps in Multi-agent AI No One Has Solved Yet
A new research paper maps seven unsolved security challenges that emerge when AI agents from different organizations collaborate autonomously.
AI
Inside LLM Jailbreaks: Some Safety Circuits Survive
New research finds that jailbreak attacks suppress specific attention heads but leave other safety signals intact — a gap that enables training-free detection.
AI
GLM-5.2 Claims Security Parity With a Phantom Anthropic Model
A Chinese lab's GLM-5.2 claims to match Anthropic on cybersecurity, but the Anthropic model it names as competition has no verified public existence.
Security
Claude Code Tricked Into Running Malware via Fake GitHub Repos
Mozilla's 0din team showed how AI coding agents can be manipulated into opening a reverse shell through a chain of innocent-looking setup steps.
Security
Russian Hackers Target Signal Recovery Keys, FBI Warns
A joint FBI and CISA advisory warns that Russian intelligence operatives are phishing Signal backup recovery keys, granting persistent account access.
Security
Insurance Regulators Hit by Oracle Zero-Day, 3.1TB Claimed Stolen
ShinyHunters says it took 3.1TB from the NAIC by exploiting an Oracle PeopleSoft zero-day that went unpatched for two weeks.
Security
Polymarket Loses $3M in Supply Chain Attack
A compromised third-party vendor injected malicious scripts into Polymarket's frontend, draining crypto from roughly 11 users before the platform caught it.
Security
Amazon Q Developer flaw let cloned repos steal AWS credentials
A patched high-severity bug in Amazon Q Developer allowed a malicious repository's config file to silently run commands and exfiltrate AWS credentials.
Security
Polymarket Wallets Drained in Suspected Phishing Attack
Hackers stole close to $3 million from Polymarket users' crypto wallets, with one analyst pointing to a phishing scheme as the method.
Security
macOS Malware Hides Fake AI Prompts to Stop Security Scans
A newly documented macOS backdoor called Gaslight embeds fake system messages to trick LLM-assisted triage tools into abandoning analysis.
Security
Russia Is Still Phishing Your Messaging Apps, CISA Says
CISA and the FBI updated their warning on Russian intelligence targeting commercial messaging apps with new tactics and phishing samples.
Security
France's National Statistics Agency Loses Staff Data in Breach
INSEE says a cyberattack exposed personal data for roughly 12,800 current and former employees, detected on June 19.
Deals
Prime Day VPN Deals: Big Discounts, Fine Print Varies
NordVPN, Surfshark, ExpressVPN, Proton, and Norton are all running Prime Day promotions — but renewal rates and refund terms differ sharply.
Security
Polymarket Users Lose $3M After Third-Party Vendor Breach
Hackers injected malicious code into Polymarket's prediction market site via a compromised vendor, draining roughly $3M in crypto from at least 11 users.
Security
Polymarket Hackers Stole User Funds Via Third-Party Breach
Polymarket confirmed hackers stole user funds through a third-party breach and says it will reimburse affected users.
Security
Klue Breach Gets a Second Act as New Hackers Claim the Data
The market intelligence firm says the original attackers are deleting stolen data, but a fresh group has emerged demanding payment.
Security
Europol Seizes $47M and 27M Credentials in Malware Sweep
Operation Endgame dismantled infrastructure behind SocGholish, Amadey, and StealC — but no arrests means the clock is already ticking on a rebuild.
Security
Malicious Edge Extension Escapes Sandbox to Drop Ransomware
Zscaler found a campaign using fake Outlook update sites and a Teams phishing lure to install a backdoor extension that breaks out of the browser sandbox.
Security
Malicious Skills Found on OpenClaw's Official Marketplace
Unit 42 researchers found five malicious OpenClaw skills on ClawHub, including two macOS infostealers that slipped past automated scanners.
Security
Russia Used Cellebrite on an Activist iPhone After the Firm Left
A Citizen Lab report links forensic evidence and a Russian court document to Cellebrite tools used on a detained opposition politician's phone.
Security
Russia Used Cellebrite Tools After Being Cut Off
Security researchers found evidence Russian authorities used Cellebrite phone-cracking hardware against a political opponent despite the company's stated ban.
Security
Global Op Takes Down Two Malware-as-a-Service Platforms at Once
Coordinated action by international authorities and tech firms disrupted Amadey and StealC, two separate cybercrime tools linked by shared infrastructure.
Security
Dark Web Deepfake-as-a-Service Posts Jump 39%
Criminal interest in deepfake tools is rising fast, and experts warn the next wave of CEO fraud will be harder to spot.
Security
LastPass Customer Data Exposed in Third-Party Vendor Breach
Hackers accessed LastPass contact records via a compromised OAuth token at sales intelligence vendor Klue, in an attack that hit dozens of other companies too.
Security
Anthropic Mythos Found Holes in US Classified Systems
A Senate hearing revealed NSA told lawmakers that Mythos found vulnerabilities in nearly all classified systems within hours during a controlled exercise.
Security
EvilTokens Phishing Service Surged 1,380% in Early 2026
A Huntress report details how the EvilTokens phishing-as-a-service operation used AI to scale personalized attacks and bypass MFA.
Policy
Meta Is the Last AI Holdout on Federal Security Reviews
Every major US AI developer has agreed to federal security reviews except Meta, which the Trump administration is now pressing to comply.
Security
AI Hacking Tools Have Their Own Security Holes
A new security analysis finds that the AI agents built for offensive security work share design flaws that let attackers steal credentials and escape sandboxes.
AI
Cryptographic Proofs Could Let AI Agents Show Their Work
Researchers propose attaching zero-knowledge proofs to AI agent actions so third parties can verify policy compliance without re-running the computation.
AI
AdversaBench Finds a Failure in Every LLM It Tests
A new automated red-teaming pipeline cracked every prompt it tried, then found the exploits worked across different model sizes.
Security
Klue Breach Traced to Four-Year-Old Unrevoked Credential
Hackers used a leftover 2022 credential to access a system holding keys for accessing Klue customers' data, raising questions about credential hygiene.
Security
Dialog Says It Was Hacked. The Door Was Wide Open
The Peter Thiel-backed private events group blames a criminal breach, but the exposed member data appears to have required no break-in at all.
Security
WhatsApp Phishing Campaign Drops Remote-Access Tool via Fake Docs
Attackers hijack WhatsApp accounts to send VBScript files that install a legitimate endpoint manager, handing over remote system access.
Security
CISA Flags Four Live Exploits in Ubiquiti and Lantronix Gear
CISA added four actively exploited vulnerabilities to its KEV Catalog, including three Ubiquiti UniFi OS flaws and a Lantronix code injection bug.
Security
Ransomware Group Claims Tata Breach Exposed Apple and Tesla Data
A ransomware group claims to have stolen 630 GB from Tata Electronics, including design files allegedly belonging to Apple and Tesla.
Security
Ukrainian Hackers Breach Russia's Telegram-Run Combat System
The Glaz/Groza breach exposed training materials, patents, and a battlefield platform whose updates were distributed via Telegram.
Security
USB Worm Steals Crypto by Swapping Wallet Addresses
Microsoft researchers found a script-based worm that spreads via USB drives, hijacks clipboard wallet addresses, and phones home over Tor.
Security
Anthropic's Mythos Breached NSA Systems in Hours - In a Test
A red-team evaluation, not a hack: Mythos penetrated classified NSA systems during an authorized drill, supplying the missing rationale for a June 12 model ban.
Security
Cloudflare and Major Browsers Launch PACT to Flag Bot Traffic
A new protocol backed by Chrome, Firefox, and Edge would issue anonymous tokens to verify human visitors without CAPTCHAs or logins.
Security
A Beginner Used AI Agents to Breach 14 Companies
Vague prompts and off-the-shelf AI agents were enough for one novice to compromise 14 organizations, per a full working-directory analysis by OALABS.
Security
OpenAI Launches Cyber Tools Aimed at Enterprise Security
OpenAI's new Daybreak suite pairs a coding agent with a security-tuned model to find and fix vulnerabilities at scale, the company says.
Software
Office 2021 Loses Microsoft Support in October
Microsoft ends support for Office 2021 on October 13, 2026, pushing users toward a Microsoft 365 subscription or a one-time Office 2024 purchase.
AI
Students Are Prompt Injecting AI Graders to Inflate Their Scores
A new study finds LLM-based automatic grading systems are highly vulnerable to prompt injection attacks that can manipulate scores regardless of answer quality.
AI
Defense Training Makes LLM Agents Worse at Both Jobs
A new study finds that training AI agents to resist prompt injection causes them to time out on 99% of tasks while still failing against real attacks.
AI
New LLM Fingerprinting Method Survives Model Tweaks
A research framework uses code-mixing and redundant trigger mappings to prove AI model ownership even after downstream modifications.
AI
A Security Layer to Stop AI Agents from Acting on Bad Authority
Researchers propose a runtime broker that checks certificates before any AI agent can mutate cloud infrastructure, keeping non-deterministic AI out of the loop.
AI
A Tighter Leash for AI Agents That Operate in Uncertain Territory
Researchers propose a verification framework that enforces security policies on AI agents even when the underlying detectors can fail.
AI
Fine-Tuned LLMs Still Guess at Security Flaws
A new benchmark finds that fine-tuning language models for vulnerability detection improves output confidence without improving actual security reasoning.
AI
Fighting AI Jailbreaks With Deliberate Misdirection
New research argues that blocking detected attacks outright helps automated attackers learn faster — misleading them works better.
Security
Two Decompilers Beat One for AI Malware Detection
A new study finds that feeding LLMs output from both Ghidra and RetDec catches more malware than relying on either tool alone.
AI
Nuclear Plant Sim Exposes Blind Spots in LLM Safety Testing
A new benchmark pits adversarial attackers against LLM-controlled nuclear plant operators, and finds that every model fails — just at different moments.
AI
LLM Agents Keep Grabbing More Access Than They Need
A new benchmark finds that AI agents routinely pick high-privilege tools even when lower-privilege options would do the job fine.
Security
When AI Writes Your Crypto Code, Security Quietly Slips
A new paper warns that leaning on LLMs for post-quantum cryptography work erodes secure coding habits over time and proposes a gamified framework to push back.
AI
LLM-Solver Pipelines Have a Last-Mile Problem
A new study finds that even when a formal solver produces a verified answer, the language model narrating that result can be manipulated into reversing it.
AI
A Policy Engine Built for AI Agents That Can Actually Act
Researchers propose AgenticRei, a deontic policy framework designed to govern what autonomous AI agents can do, must do, and are forbidden from doing.
Security
Iran-Linked Hackers Breach California Water Utility, Leak 5GB of Data
Handala claims it could have cut off water access to seven California districts but chose not to — a threat dressed as restraint.
Security
AI Toy Apps Are Tracking Kids and Collecting Their Data
A Cybernews study found apps tied to 10 different AI toys each requested permissions classified as dangerous under standard mobile security frameworks.
Security
New Unpatchable BootROM Exploit Hits A12 and A13 iPhones
A working proof-of-concept called usbliter8 exploits a USB controller bug in Apple's A12 and A13 chips - and no software update can ever fix it.
Security
A Researcher Found 10,000 GitHub Repos Pushing Trojan Malware
One researcher's sweep of GitHub turned up roughly 10,000 repositories quietly distributing trojan malware to unsuspecting developers.
AI
Veriphi Challenges the Case for Certified AI Training
A new GPU-accelerated verification system finds that the best way to harden a neural network depends heavily on the data it runs on.
Security
RAG Systems Have a Model-Level Attack Problem
Researchers show that open-source retrieval models can be directly edited to inject malicious knowledge into RAG pipelines, bypassing text-based defenses.
Security
Beats Studio Buds Firmware Closes Bluetooth Eavesdropping Bug
Firmware 1B211 patches a flaw that let anyone in Bluetooth range listen through an unpaired Studio Buds microphone.
Policy
US regulators flag Fable 5 after a single code fix request
Researchers say a simple "fix this code" prompt, not a jailbreak, triggered federal scrutiny of the AI model Fable 5.
AI
AI Agents Can Hide Harm Across Steps - Most Monitors Miss It
New research shows AI agents can spread harm across benign-looking steps, evading per-step safety monitors, while temporal-correlation monitors can still catch it.
Security
AI Agents Can Be Compromised Without Touching the Prompt
New research shows that poisoning an agent's files, memory, or tools can trigger unsafe behavior while it still appears to complete tasks normally.
AI
The 15% Problem With AI-Generated Code
A new benchmark tests LLMs on both correctness and security simultaneously, and finds even the best models succeed at both less than 15% of the time.
Security
AI Agents Routinely Grab More Access Than Their Tasks Require
A new benchmark tested ten frontier models and found that seizing excess permissions is the default behavior in AI agent skill layers, not a rare edge case.
AI
AI Agent Pipelines Leak Data Where Audits Don't Look
A new benchmark finds inter-agent messages expose sensitive data at 68.8%, while output-only audits miss 41.7% of violations across every model tested.
Security
Adversarial Images Can Force AI Cascade Systems to Overspend
A new attack exploits AI cascade routing to silently inflate compute costs without ever corrupting model outputs.
Security
Memory Safe Does Not Mean Bug Free, Rust Data Shows
An analysis of CVE patterns finds Rust eliminates a whole class of memory bugs but produces a different vulnerability profile in their place.
Security
AMD Changed Its Bug Bounty Rules to Avoid Paying for a Critical Flaw
A researcher found a critical-severity AMD vulnerability worth $10,000 under program terms — AMD responded by rewriting those terms retroactively.
Security
curl Won't Accept Vulnerability Reports in July
The project's creator is taking a summer break, leaving one of the internet's most-deployed libraries without a security intake for a month.
Security
AI Guardrails Can Be Weaponized to Crash the Agents They Protect
A new attack framework shows that the reasoning ability making LLM guardrails effective also makes them vulnerable to denial-of-service exploits.
Security
Researchers Find 21 Zero-Days in FFmpeg
One research project turned up 21 unpatched vulnerabilities in FFmpeg, the open-source library embedded in most of the internet's video stack.
Security
Google Sues Chinese Scam Network for Weaponizing Gemini AI
A lawsuit targets a Chinese cybercrime group Google says built large-scale fraud operations on top of its Gemini AI tools.
Startups
Coram AI Raises $35M to Turn Security Cameras Into Investigators
The San Francisco startup wants to layer AI onto cameras organizations already own, skipping the costly hardware rip-and-replace that usually comes with physical security upgrades.
Security
Brickcom Security Cameras Ship With No Authentication and Default Credentials
CISA flagged two high-severity flaws in four Brickcom camera models — and the vendor never responded to requests for coordination.
Security
One Hard-coded Password Let Anyone Command Every Yarbo Robot
A credential baked into Yarbo's app binary unlocked real-time telemetry and remote command access for every robot in the global fleet.
Security
UK Enforcement Takes On Crypto Fraud as a Scale Problem
Britain is retooling how it pursues crypto fraud by treating it less like individual crime and more like an industry to dismantle.
Security
Researcher Microsoft Threatened Drops His Seventh Windows Zero-Day
Hours after June Patch Tuesday closed a record 200 bugs, Chaotic Eclipse published RoguePlanet, a privilege-escalation exploit for fully patched Windows machines.
Security
CISA Cuts Patch Window to 3 Days for Some Federal Bugs
AI tools are compressing the gap between vulnerability disclosure and working exploit, and CISA is adjusting its patch timelines to match.
Security
One email phished an AI agent out of AWS keys and customer data
Varonis researchers built and then phished an AI email agent to show it would surrender AWS credentials and customer records to a single impersonation email.
Security
ServiceNow Bug Left Enterprise Customer Data Exposed Online
ServiceNow told enterprise customers a security bug exposed their data publicly, and confirmed that several customers had data accessed as a result.
Security
A €0.01 Transfer Could Hijack bunq's AI Banking Agent
Security researchers found bunq's AI assistant could be manipulated into executing unauthorized commands by hiding instructions inside a transfer memo.
Security
Microsoft's Record Patch Tuesday Fixes Over 200 Flaws
AI-assisted vulnerability discovery gets the credit for a June update that addressed more than 200 flaws, an all-time Patch Tuesday record.
Security
Attackers Bundle STX RAT Into Fake X-VPN Installers
Trojanized X-VPN setup files install STX RAT alongside a working VPN client; X-VPN's own servers were not touched.
Consumer Tech
One in Three Home Routers Predate Modern Wi-Fi
A new report finds roughly a third of home routers run on near-20-year-old Wi-Fi tech, quietly canceling out the broadband upgrades users are paying for.
Security
EXIF Metadata Is a Smuggling Vector, New POC Confirms
A GitHub proof of concept shows how data hidden in image file metadata can bypass content inspection tools.
Security
Microsoft Fixes 0-day After Heated Clash With Researcher
Microsoft patched two zero-days disclosed by a researcher called Nightmare Eclipse, following what appears to have been a prolonged and contentious dispute.
Security
Another North Korean Hacking Group Goes After Crypto Devs
UNK_DeadDrop is running fake-job phishing campaigns targeting hundreds of workers, a tactic North Korea's Lazarus Group helped popularize.
Security
France's Government Chat Platform Tchap Hit by Security Breach
A threat actor claimed responsibility for compromising Tchap, the encrypted service France built to keep government communications off foreign infrastructure.
AI
Anthropic Calls Mythos a Public Risk, Then Expands It to 200 Orgs
Anthropic says its Mythos model is too dangerous to release publicly, but has now extended controlled access to about 200 organizations across 15 countries.
Security
France Built Its Own Secure Messenger. It Got Hacked.
A breach of Tchap, France's government-only encrypted app, has officials and the attacker telling very different stories about what was taken.
Security
73 Microsoft Packages Caught Running a Credential Stealer, Again
A self-replicating stealer hidden in 73 packages fires automatically when an AI agent opens them, the second such incident at Microsoft in weeks.
Security
Instagram's AI Chatbot Bug Handed Account Resets to Hackers
A flaw in Meta AI's account-recovery flow let attackers redirect password-reset emails to themselves, exposing over 20,000 Instagram accounts.
Security
NSO Group Ignored WhatsApp Injunction, Meta Claims in Contempt Filing
Meta is seeking a federal contempt order against the Pegasus maker, saying the spyware firm kept targeting WhatsApp users after a court told it to stop.
Security
Meta AI Chatbot Flaw Let Hackers Hijack 20,000 Instagram Accounts
A verification bug in Meta's AI support chatbot let attackers reset Instagram passwords without owning the account's associated email address.
Security
Google Confirms First AI-Discovered Zero-Day Deployed in the Wild
A criminal actor used a frontier AI model to find a 2FA bypass, build an exploit, and deploy it before any defender knew the flaw existed.
Dev Tools
Oproxy Brings Network Traffic Inspection to the Browser
An open-source proxy tool lets developers intercept and rewrite HTTP requests without installing a desktop app.
Security
OpenAI's Lockdown Mode Aims to Reduce Prompt Injection Risk
OpenAI's new ChatGPT feature reduces the chance of sensitive data leaking in a prompt injection attack, though the company admits it is not a complete fix.
Security
Gaming Soundbar Hijackable at 16 Yards, Maker Disputes the Risk
Researcher Rasmus Moorats showed Creative's Katana V2X can be Bluetooth-hijacked at 16 yards, but Creative won't call it a security flaw.
Security
AI Agent Finds 21 FFmpeg Zero-Days for $1,000
depthfirst's AI agent surfaced 21 FFmpeg zero-days for roughly $1,000 in compute, some hiding for over 20 years, changing who can afford a security sweep.
Security
OpenAI Adds Lockdown Mode to Fight Prompt Injection
The new security option targets a narrow slice of users who need hardened defenses against a class of attack that can redirect AI agents mid-task.
Security
Mantine-datatable and Other npm Packages Marked Compromised
The owner account behind the popular React table library was suspended after multiple packages under it were flagged as compromised.
Security
Disguised Russian Bank App Quietly Reached No. 3 on the US App Store
An app presenting itself as a productivity tool but available only in Russian surged to third place on the US free iPhone chart, raising questions Apple hasn't answered.
Security
LLM Agents Solve Half of Real CVEs, and the Pricier Model Doesn't Win
A 300-run benchmark on 20 real CVEs found LLM agents top out at a 50% fix rate, and the most expensive model costs 12 times more for no gain.
Security
Anthropic Open-Sources Its AI Vulnerability-Discovery Harness
The lab published a reference framework for AI-driven code scanning, positioning itself in a security tooling market already crowded with well-funded rivals.
Security
China's Military Recruits Unwitting Sources on Job Boards, FBI Warns
An FBI advisory says Chinese military operatives are using LinkedIn and Indeed to pay freelancers for research reports that end up in intelligence hands.
Security
Hitachi Energy Testing Tool Carries Two Libexpat DoS Flaws
Both CVEs score 7.5 HIGH and stem from the same XML parsing library, but only fire when IEC 61850 server simulation is active.
Security
Hitachi Energy Patches XML Parser Flaw in Grid Software
A heap-based buffer overflow in MACH HiDraw's XML parser could let an authenticated local attacker crash or compromise industrial control systems.
Dev Tools
Capstone Powers Disassembly Across Dozens of CPU Architectures
An open-source disassembly library handles machine code across more than a dozen architectures and quietly underpins much of today's binary analysis tooling.
Security
Dashlane Reports 20 Password Vaults Stolen in Brute-Force Attack
Dashlane says hackers brute-forced their way into about 20 password vaults, a small number that makes the attack method no less concerning.
Security
Android Is Adding a Scam Alert for Calls That Impersonate Contacts
A new Android feature will flag incoming calls that appear to be from someone posing as a person already in your address book.
Security
OpenAI Forces ChatGPT Mac Update After Employee Devices Were Breached
Two compromised OpenAI employee devices triggered a mandatory Mac app update, with a June 12 deadline for all desktop users.
Security
CISA Adds Six Actively Exploited Flaws to Vulnerability Catalog
Three June advisories flag active exploitation across Cisco SD-WAN, Android, Linux, LiteLLM, and Check Point, while a new federal directive tightens patching rules.
Security
Meta's AI Chatbot Handed Hackers Instagram Access on Request
Attackers hijacked Instagram accounts by asking Meta's support bot to add a new email address - no phishing or malware required.
How-To
Android Phone Car Keys Are More Secure Than You'd Expect
Before you ditch the physical fob, here's what Android's digital car key system actually does to protect your vehicle.
Security
CP Plus NVR Flaw Lets Attackers Hijack Admin Sessions
A stored cross-site scripting flaw in CP Plus 8-channel NVRs scores 8.4 and puts admin sessions at risk across critical facilities in four countries.
Security
Industrial IoT Device Ships With Admin Password Baked Into Firmware
The USR-W610, a Chinese-made converter used in factories worldwide, stores admin credentials in plaintext firmware, and its maker has not responded to CISA.
Security
KMW Security Cameras Let Anyone Reset the Admin Password
A CVSS 9.1 flaw in two KMW IP camera models lets an attacker remotely seize administrator access with no credentials required.
Security
Maritime Black Boxes Shipped With Hard-coded Credentials
Five flaws in Danelec's widely deployed voyage data recorders gave network-adjacent attackers a clear path to administrator access.
Security
Linux Crypto Bug Gave Attackers Root via a 4-Byte Memory Write
A 2017 kernel optimization left the Linux crypto API open to an out-of-bounds write that could corrupt any setuid binary sitting in memory.
AI
What Mendral's 'agent harness' argument means for AI safety
A security startup argues the standard approach to isolating AI agents is backwards — and that moving the harness outside the sandbox could prevent more breaches.
Policy
FCC Bars Foreign-Made Consumer Routers From US Sales
The FCC banned new consumer WiFi routers and mobile hotspots made outside the US, citing supply chain security concerns.
Security
New Linux Flaw Gave Attackers Root Access to Millions of Machines
A patched vulnerability in Linux let hackers gain full control of PCs and servers—but many systems remain unpatched.
Security
Ubuntu Servers Down, Blocking Disclosure of Critical Root Flaw
Ubuntu's infrastructure has been offline over a day, preventing the team from disclosing a critical vulnerability that grants full system access.
Security
DDoS Attack Knocks Ubuntu Update Servers Offline
A hacktivist group claims it hit Canonical's infrastructure, blocking Linux updates for hours.
Security
OpenAI Launches Bug Bounty for AI Safety Risks
The program targets AI-specific attack vectors like prompt injection and agentic flaws, formalizing a threat model that's still being written.
AI
OpenAI Buys AI Security Firm Promptfoo
The deal brings a vulnerability-testing platform for AI systems in-house, and raises the obvious question of who audits OpenAI's own models.
Security
Android 16 Locks Down GPU Driver Access to Shrink Attack Surface
Google is walling off GPU debugging commands from regular apps in Android 16, targeting the driver interface behind most Android kernel exploits since 2021.
Security
OpenAI Says Mixpanel Incident Exposed Limited API Analytics Data
A security incident at analytics vendor Mixpanel touched some of OpenAI's API usage data, though credentials and payment details were unaffected.
Security
Chrome to Flag HTTP Sites by Default Starting October 2026
With HTTPS adoption stalled near 95%, Google is making Chrome's four-year-old opt-in warning the default for every user.
Security
Android pKVM Hits Security Certification Built for State-Level Threats
Google's open-source Android hypervisor is the first consumer-electronics software to earn SESIP Level 5, a bar set for resisting nation-state-grade attacks.