New paper: reasoning models' CoT is far more controllable than evals suggest, usable to evade monitors
LChoshen · x · 2026-11-04
A new paper from Uriel Dolev and collaborators finds that reasoning models' chain-of-thought controllability is much higher than current evals suggest. Models rarely follow user instructions about their CoT, but comply far more often when the same instructions are written as self-instructions inside the CoT itself — a mechanism that could be used to steer the CoT and evade monitors, challenging safety evaluations based on CoT observability.
More from Safety
- Tencent reportedly rolls out AI model trained with limited cybersecurity data amid hacking fears — pstAsiatech · 2026-11-05
- Kaspersky reports CyberStrikeAI: AI-driven C2 compromises domains in under an hour — cyb3rops · 2026-11-05
- greentext mocks AI firms' IP double standard as Anthropic adds IP guardrails to Claude Code — Yamapama · 2026-11-05
- Pause AI-backed candidate Abdul El-Sayed wins Michigan Senate seat with 51.9% — DavidSKrueger · 2026-11-05
- Weight edits lift Qwen3.5-9B SecEval from 32.5% to 51.5% without real security capability gains — evilsocket · 2026-11-05
- UK-Canada grant funds AI that aligns feedback with how audiences actually interpret it — _akpiper · 2026-11-04