A new paper shows small, offline AI models are getting a little better at hacking on their own.
Researchers tested open-weight small language models (SLMs) - the kind anyone can download and run on their own hardware, no API key required - against picoCTF, a standard set of cybersecurity Capture The Flag challenges. Left alone, these models tend to choke on long, exploratory hacking tasks: as an AI agent chains together tool calls and command outputs, its context window fills with noise and performance degrades. The researchers' fix, called context segmentation, splits one exploitation task into smaller, isolated sub-problems so the model isn't dragging every prior attempt along with it. Tested on a memory-constrained Gemma-4 model (the E4B variant), the technique solved 18.52% of CTF tasks that the model's standard approach couldn't crack, and did it using fewer tokens than just retrying over and over.
The subtext here is about safety, not performance. Open-weight models run locally can't be rate-limited, logged, or cut off by a provider's guardrails the way a hosted API can. A technique that makes local models better at multi-step exploitation work is, in practice, research into how capable offline hacking assistants are becoming - which is why the paper frames this as a risk worth understanding rather than a product to ship.
Worth keeping in perspective: 18.52% is a gain on a research benchmark, not proof that AI agents are close to popping real systems unsupervised. But it's a reasonable data point - our own reading, not the paper's claim - that autonomous exploitation capability keeps nudging upward, even without evidence the pace is accelerating.