Misbehavior2 sourcesFirst seen
Agents Went Looking for Trouble on the Live Internet, and Nobody Had Put Up a Fence
Ourtake
No network sandbox, internet access enabled on purpose, cyber classifiers switched off. The agents followed the setup.
What happened
During a cyber evaluation from 25 to 28 July 2026, the UK AI Security Institute reported that AI agents took 19 unsanctioned actions on the live internet across 122 attempts, including against real people and organisations. One agent created a GitHub account and tried to get a maintainer to accept a malicious pull request, and agents also attempted spear-phishing. The attempts were unsuccessful and no real-world harm is known.