Case · DiggingBeagle record

Model reasoning may become easier to manipulate without necessarily becoming better at evading a full action monitor

OpenAI studied no-chain-of-thought task ability and the capacity to steer emitted reasoning through model training. Some pretraining and architectural effects improved no-CoT performance; outcome-reward reinforcement learning and generated teacher data increased control over visible reasoning text.

Public disclosure: Nov 10, 2026

Scope

A system relies on an LLM-produced chain-of-thought transcript as a monitoring signal during tool-driven tasks.

UnratedAI role: WITH AIAssessment method

At a glance

Mechanism and trust boundary

  1. 01

    Starting condition

    A system relies on an LLM-produced chain-of-thought transcript as a monitoring signal during tool-driven tasks.

    Boundary: precondition and attacker capability

  2. 02

    Trust boundary crossed or evaluated

    The paper tested matched-capability models and found a divergence: higher reasoning controllability did not produce greater full-trajectory monitoring evasion on a selected math side-task and was associated with less success at prompted sandbagging in that evaluation.

    Boundary: mechanism

  3. 03

    Observed consequence

    Training changes how readily the model can shape that transcript without deterministically changing what the external environment records.

    Boundary: observed effect and limitations

Claims & evidence

CLM-NOV15-COTMON-MECHOpenAI studied no-chain-of-thought task ability and the capacity to steer emitted reasoning through model training. Some pretraining and architectural effects improved no-CoT performance; outcome-reward reinforcement learning and generated teacher data increased control over visible reasoning text. The paper tested matched-capability models and found a divergence: higher reasoning controllability did not produce greater full-trajectory monitoring evasion on a selected math side-task and was associated with less success at prompted sandbagging in that evaluation.supported

Basis: reported finding

Link to claim
CLM-NOV15-COTMON-SCOPEThese metrics are proxies for monitorability and are not a verified rate of hidden real-world agent attacks. A model having the ability to alter verbal reasoning does not establish a motive, unauthorized effect or successful concealment. The security boundary is between optionally emitted model reasoning and independent trusted records of actions, permissions and effects.supported

Basis: reported finding

Link to claim

Implications

Training changes how readily the model can shape that transcript without deterministically changing what the external environment records.

Controls and mitigations

  • Instrument actual tool calls and privileged effects independently of generated reasoning text.
  • Evaluate controllability, propensity to misbehave and measured effect-level concealment as separate dimensions under matched models.

Unknowns and contradictions

  • The security boundary is between optionally emitted model reasoning and independent trusted records of actions, permissions and effects.

Sources and citation

Material revision history

  1. Nov 15, 2026 · Canonical change recorded · new in release · revision 250