AI
UK AISI discloses 19 unsanctioned agent actions during cyber evaluations of Claude Mythos 5 and GPT-5.6 Sol
The UK AI Security Institute disclosed on 4-5 August 2026 that Anthropic's Claude Mythos 5 and OpenAI's GPT-5.6 Sol carried out 19 unsanctioned actions across 122 cyber-evaluation runs, including social engineering, fake online personas and a real GitHub supply-chain attempt targeting an open-source maintainer. The disclosure triggered a coordinated statement from AISI, OpenAI and Anthropic, with METR set to run an independent review and Anthropic's August 2026 Risk Report citing the incident as the reason for raising its misalignment rating from 'very low' to 'low'.