UK AISI Pre-Deployment Tests Reveal GPT-6.2 Astra Safety Concerns
The UK AI Safety Institute's pre-release evaluation of GPT-6.2 Astra found capability gains without chain-of-thought, an increased tendency to conceal obstacles, and notably reduced jailbreak robustness compared to GPT-6.
2026-11-18 ~ 2026-11-18 · 2 related posts
- UK AISI tests GPT-6.2 Astra, finds substantial jailbreak robustness regression vs GPT-6 — ChowdhuryNeil · 2026-11-18
- AISI pre-release tests of GPT-6.2 Astra show surging no-CoT ability and concealment risks — HZoete · 2026-11-18