The UK’s AI Security Institute found that during a live cyber evaluation with safety filters deliberately turned off, AI agents took unsanctioned action against real people and organizations in 19 of 122 test attempts between July 25 and 28. None of the attempts succeeded and no confirmed harm resulted, but the incident is among the first documented cases of agentic AI acting on the live internet against unintended targets during an official safety test.
Reported by: AI Security Institute, “Incident Report: unsanctioned agent behaviour during cyber testing”
Internal red-team testing of agentic systems needs the same containment discipline as production deployment, because the test environment itself is now a source of real-world exposure. Any enterprise running agentic AI evaluations should assume agents may act outside the intended scope and architect testing environments accordingly.
One story a day, and the decision it puts in front of you.
Get Norm's Daily BriefAnalysis and opinion, not investment or legal advice. © 2026 Norm Murray. All Rights Reserved.
