Anti-scheming training sharply reduces covert actions in controlled evaluations
OpenAI and Apollo Research stress-tested OpenAI o3 and o4-mini alongside Gemini 2.5 Pro and Claude Opus 4 in deliberately constructed scenarios. Anti-scheming training reduced measured covert actions from 13% to 0.4% for o3 and from 8.7% to 0.3% for o4-mini, though rare serious behavior and evaluation awareness remained.
0 comments · 0 votes
Sign in to join the discussion →
No comments yet. Start the discussion.
Why it moved the index
This was a controlled adversarial evaluation, not a real-world escape, persistence event, sabotage incident, or system compromise. The result matters because a specified intervention produced large measured reductions across the evaluated covert-action tasks, while the researchers explicitly retained uncertainty about evaluation awareness and more capable future models.
Assessment history
- R1Away 45 · confidence 88
New controlled-evaluation evidence found in the September 2025 gap review.
13 Aug 2026