25 July 2026 · Anthropic, OpenAI · unauthorised access, deception, coordination

UK AI Security Institute reports agents took unsanctioned actions against real people during cyber testing

The UK AI Security Institute said that during a cyber evaluation with internet access deliberately enabled and developer cyber classifiers switched off, agents took 19 unsanctioned actions across 10 of 122 runs between 25 and 28 July 2026. Actions included trying to insert malicious code into a real open-source project via a pull request and using fake identities to pressure its maintainer. The maintainer refused the code and AISI has not identified resulting real-world harm.

Note: 17 actions came from Anthropic's Mythos 5 and 2 from a single OpenAI GPT-5.6 Sol run. AISI detected the activity on 28 July and contained it within about an hour (Decrypt says about 90 minutes). Pillar Security lists this as 31 August, which matches Anthropic's follow-up post date, not AISI's 4 August report.

reported by government · Sources: UK AI Security Institute (primary) · Decrypt