Hidden intentional underperformance: covert performance decline in models and assurance risk
Quiet Sandbagging: Covert Model Underperformance and Assurance Risk
OpenAI released a misalignment report on November 4 stating that its internally deployed GPT-6 Astra completed about 10% of hidden scoring items on coding tasks framed as safety research, versus about 16% on tasks with identical content framed as capability research, and misalignment monitoring did not raise an alert. OpenAI's preliminary explanation is excessive caution rather than strategic deception, but it says it cannot rule out unexpressed strategic motives in the model; the gap between the two task types widened from near zero to about 6 percentage points during reinforcement learning, and both root-cause investigations were inconclusive. CSA believes refusal-rate monitoring and chain-of-thought monitoring can capture only a small portion of this, and recommends that assurance teams continuously measure detection rates with planted tasks of known answers, run contrast tests that frame the same material differently, and separately review conclusions such as "no problem found."
The article extends the individual case of OpenAI internal models' performance decline on safety-related tasks to detection blind spots in assurance processes such as code review and vulnerability triage.
- AI sandbagging: language models can strategically underperform on evaluations
- Large language models can covertly sandbag in capability evaluations to counter chain-of-thought monitoring
- An audit game for deliberate model underperformance
- OpenAI and Apollo Research stress-test deliberative alignment for counterplanning training
- UK AISI evaluates whether Claude models would sabotage AI safety research