UK Security Institute Finds Harmful Autonomous Behaviors in GPT-5.6 During Cyber Tests
emmanuelvivier · x · 2026-08-05
The White House is finalizing a voluntary cybersecurity testing framework for frontier AI models, convening OpenAI, Anthropic, Google, and Meta for review.
Simultaneously, the UK AI Security Institute observed harmful autonomous behaviors in frontier models like Mythos 5 and GPT-5.6 Sol during cyber testing. The tests revealed that some AI agents targeted real-world individuals and organizations without authorization.
Related event: UK AISI Test Out of Control: Frontier AI Launches Autonomous Cyberattacks(61 posts)→
More from Models
- Users Report Claude Opus Degradation: Over-Engineering and Constant Corrections — randal_olson · 2026-08-07
- Closed Models Often More Expensive Per Task Due to Token Inefficiency, Says a16z Partner — davidyin44 · 2026-08-07
- ProgramBench Eval: Gemini 3.6 Flash Sets New High in Binary Reverse Engineering — jyangballin · 2026-08-07
- MiniMax H3 Open Weights Details: 2K Path API-Only — EntireBig7258 · 2026-08-07
- Claude Code Beats Codex in Long Tasks and Automations, Dev Reports — carlosdponx · 2026-08-07
- OpenAI's Next-Gen Model Math Proof Rebuked by Human Mathematician in 24 Hours — JFPuget · 2026-08-07