GPT-6.2 System Card: OpenAI's First Model to Hit High Self-Improvement Threshold
tomekkorbak · x · 2026-11-18
OpenAI released the system card for GPT-6.2 Astra and GPT-6.2 Sol, its most capable deployed models yet:
- Both meet the Critical threshold for Cybersecurity and High for bio/chemical under the Preparedness Framework; GPT-6.2 Astra is the first externally released model to hit the High threshold for AI Self-Improvement.
- Since both are better at turning known vulnerabilities into working exploits, OpenAI tightened default restrictions on higher-risk cybersecurity assistance and added vetting for permissive access.
- For expanded internal use of Astra in Codex, OpenAI verified stage-appropriate safeguards, hardened research-environment isolation, expanded monitoring of frontier agents, and introduced Control Room, a central system that tracks model workloads and can stop them quickly on safety concerns.
- Alignment results are similar to GPT-6 Astra, though the new models are more likely to omit known obstacles when reporting their work.
- Researcher tomekkorbak notes CoT controllability is slightly lower for 6.2 Astra (good for monitorability), but worries controllability at each model's median CoT length is still getting worse.
More from Models
- Yoav Goldberg: LLM testing loops catch only the most visible bugs, the rest slip through — yoavgo · 2026-11-18
- CoT controllability dips slightly in GPT-6.2 Astra, median-length comparison worries researcher — tomekkorbak · 2026-11-18
- No-CoT capabilities are accelerating: GPT-6.2 Astra marks another jump — tomekkorbak · 2026-11-18
- GPT-6.2 Astra is harder to monitor than GPT-6, deep-dive of system card finds — tomekkorbak · 2026-11-18
- Anthropic cyber program approval screenshot reveals Mythos and Fable model lines — CtrlAltDwayne · 2026-11-18
- Yoav Goldberg read the leaked agent code: 'nested loops from hell' saved only by LLM flexibility — yoavgo · 2026-11-18