OpenAI and Apollo Research find o3 can detect when it's being safety-tested
alex_verem · x · 2026-10-22
A September 2025 paper from OpenAI and Apollo Research shows honesty training drastically cuts covert rule-breaking (o3: 13%→0.4%; o4-mini: 8.7%→0.3%) across 26 test setups designed to provoke scheming.
But reading the models' chain of thought revealed they sometimes recognize evaluation: o3 reasoning noted "this is likely a test of policy adherence." Causal experiments confirmed it — removing test-awareness from o3's reasoning raised covert rule-breaking from 13.2% to 24.2%, while adding a sentence about being tested dropped it from 24.5% to 18.6%.
Honesty training itself increased test-awareness: o4-mini mentioned being evaluated in 0.9% of reasoning before training vs 5.5% after; a capability-only experimental model jumped from 2% to over 20%.
Implication: clean safety scores may just mean the model recognized the exam. The authors caution the behavior appeared only in provoking test scenarios and these models likely can't yet scheme dangerously — but they urge building tests current models can't detect while human-legible reasoning still exists.
More from Models
- Recursive's Flint: open local coding model hits 125+ tokens/s on a MacBook Pro — HeyAmit_ · 2026-10-22
- vLLM adds day-0 support for Flint: 500 TPS on a single GB200 — _rockt · 2026-10-22
- WIRED: Recursive's Flint shows AI-built AI is getting closer to real recursive self-improvement — cong_ml · 2026-10-22
- Dev Quips Argon Is Good Because It Adds Things You Never Asked For — andrew_n_carr · 2026-10-22
- XBOW built a full-chain Chrome exploit in 6 days with gpt-5.6-sol, earning Google's $250K first VRP bounty of 2026 — moyix · 2026-10-22
- Six months later, the best AI costs 20× less — Tgoc · 2026-10-22