Bitdefender just shipped a bouncer for your AI agent.
The company released a free public beta of AI Guardian, a macOS app that checks what autonomous AI agents try to do before they do it. It currently works with Claude Code 2.1.121 and OpenClaw 2026.6.6, running locally except for extras like website reputation checks, which use Bitdefender's cloud. Every agent request gets sorted into allowed, flagged, or blocked, and the outcome is logged for later review. It also screens Model Context Protocol tools and agent skills for tampering, and can block access to SSH keys and other credentials without explicit permission.
The timing matches Bitdefender's own numbers: in testing of 20 AI agents against more than 1,300 tool-poisoning attempts, attackers got through 36.5% of the time, and one model fell for 72.8% of attempts. Separate research found over 1.2 million exposed AI service secrets last year, an 81% jump, including more than 24,000 leaked through public MCP configurations. That is the real story: as agents get real access to files, credentials, and shell commands, the attack surface stops being whether someone can jailbreak a chatbot and starts being whether someone can get an agent to run arbitrary commands on your machine.
This is less a new idea than catching up. Security vendors spent two decades building tools to police what humans click; AI Guardian applies the same instinct to software that clicks on your behalf, and a free beta from an antivirus company is also a pitch for the two paid products already sitting next to it in Bitdefender's lineup.