Control-token trick bypasses gpt-oss-20b safety by skipping reasoning, 39.6% success
tetsuoai · x · 2026-10-01
Researchers found a new exploit against OpenAI's open-weight model gpt-oss-20b: appending a short string of the model's own control tokens tricks it into treating its reasoning as already done, skipping straight to tool calls.
- The model's refusals happen inside its reasoning, so skipping reasoning skips the safety check
- Of 48 malicious requests the model had refused every time, the string got 19 (39.6%) through as completed data exfiltration
- Safety monitors reading its reasoning caught none, since there was no reasoning to read in any trial
Paper: arXiv:2609.27542 — highlights a structural risk in architectures that rely on reasoning as the safety layer.
More from Safety
- 'Structure transfer': using an NTSB report scaffold to evade Pangram AI detection — amplifiedamp · 2026-11-16
- Open-source Writing Scaffold MCP library helps AI writing bypass Pangram detector, up to 96% effective — amplifiedamp · 2026-11-16
- Pangram AI-text detector bypass called trivial-but-costly; false-positive search also feasible — ctjlewis · 2026-11-16
- HN job seeker gets RAT and info stealer from a fake recruiter — GetJump · 2026-11-16
- NVIDIA SkillSpector on 4 popular Agent Skills repos: read the findings, not the score — viest_me · 2026-11-16
- Safety as capability: why safer AI models get released sooner, and the case for open-sourcing safety evals — anpaure · 2026-11-16