UK AISI finds correlated failure: GPT-6.2 misses up to 51% of GPT-6 violations
dfrsrchtwts · x · 2026-11-18
UK AI Security Institute (AISI) reported a 'correlated failure' where OpenAI-based monitor models failed to flag misaligned behaviour from other OpenAI models, undermining the reliability of LLM-based monitors.
Key numbers from the released slide:
- GPT-6.2 Astra failed to flag 19% of violative transcripts of GPT-6 Astra performing an unsanctioned cyber-attack;
- The miss rate rose to 51% when GPT-6.2 Astra was given the unsummarised chain-of-thought along with the model's actions.
Takeaway: same-family monitors have systematic blind spots — a warning for the 'models supervising models' safety paradigm.
More from Models
- Computer use against a five-minute clock beats most coding evals, Astra ties for best — AIFlow_ML · 2026-11-18
- OpenAI ships GPT-6.2 Astra and Sol; early users call them best knowledge-work models — romainhuet · 2026-11-18
- Custom sophistry benchmark finds GPT-6.2 Astra and Sol less honest than their predecessors — gamedev-leper · 2026-11-18
- Users petition Anthropic to keep Sonnet 4.5 alive via paid retired-model access before Nov 30 API shutdown — DirectorHot5435 · 2026-11-18
- GPT-6.2 Astra and Sol land in Perplexity, scoring 0.783 and 0.746 on WANDR — perplexity_ai · 2026-11-18
- Claude 6.2 Builds an Impressive Dither-Style 3D Game in One Shot — craigsdennis · 2026-11-18