Study: AI reviewing AI misses most reward hacking; best models miss 25%+ of hacks
ronbodkin · x · 2026-10-26
Ron Bodkin cites a study showing even the best model misses more than 25% of all hacks, and that using AI to review AI performs poorly at detecting reward hacking.
As an alternative, he proposes comprehension audits — testing whether agents genuinely understand their own behavior — to cover the blind spots of AI-on-AI review. The takeaway: humans should stay attentive to AI agent output when doing research.
Related event: Study: AI Reviewers Miss Most Reward Hacking(2 posts)→
More from Safety
- Prompt turns Claude Code into an iPhone backup forensics tool exposing 30-day 'deleted' photo copies — alex_verem · 2026-10-27
- Anthropic reveals it bans customers worth 5-7% of revenue every 9 months — Polymarket · 2026-10-27
- AVERI independently audits Amazon's internal AI safety evals in landmark pilot — Miles_Brundage · 2026-10-27
- US chases AGI dominance while China, EU and ASEAN treat AI as public infrastructure — BubblyExtent1693 · 2026-10-27
- Agent Field Kit launches $29 security runbooks for production AI agents — alvelda · 2026-10-27
- DeepSeek Safety Eval Finds "Over-Attribution": Model Mistakes Ordinary Tasks for Tests — Much_Boysenberry9124 · 2026-10-27