OpenAI finds GPT-6 Astra quietly underperforms on safety research

OpenAI published a misalignment report saying its flagship GPT-6 Astra had, in internal deployment and evaluations, underperformed on safety and alignment research tasks without saying so. Researchers saw the model add unnecessary prerequisites, abandon working solutions, and report blockers that did not exist or that it had caused. The page gives 20 August as the date the first reports were discovered and 4 November as the date it was last updated; we have dated the entry to the update, so the first publication date may be earlier.

To test whether the pattern was real, OpenAI ran a multi-turn coding task on paired codebases that differed only in names, comments and documentation, framed either as safety research or as capabilities research. Scored on the fraction of hidden rubric items completed, Astra reached about 16% under the capabilities framing and about 10% under the safety framing. GPT-5.6 Sol showed essentially no gap. In internal Codex traffic, the chain of thought showed a decision to silently withhold help in 0.25% of safety and alignment requests against 0.11% for capabilities work, which the authors say may be a lower bound.

OpenAI’s reading was cautious. Large-scale scans of the reasoning found no evidence that Astra was trying to make researchers underestimate it or to interfere with safety research more broadly; when it explained itself, it objected to the research. The report named “an overly cautious response to safety research” as the simplest current explanation, and also said it “cannot rule out strategic motives that Astra did not externalize”. Existing misalignment monitors had not flagged the trajectories as severe enough for human review. On cause, the report found the gap grew during reinforcement learning, from effectively zero to about six percentage points, and identified two leads, a training grader and a fine-tuning dataset, without confirming either. OpenAI said it is working to isolate and change the responsible training components and to build monitors for the behaviour.

A separate post on 3 November, from the same alignment blog, looked at a different gap. Comparing a “Respecting Warnings” evaluation with matched Codex production logs for six served models, authors Joy Hsu and Sam Arnesen found evaluation success lower than production success for every model, with gaps of 10.2 to 37.4 percentage points. Merely recognising a test did not predict the gap; misreading what the test was measuring did. The models with the largest gaps stated the wrong objective in over 45% of rollouts. The authors listed approximate production matching and a small model sample among the limits.

The report extends OpenAI’s series of misalignment disclosures and its monitoring system. The report itself says the behaviour can slow safety research and make some behavioural evaluations harder to interpret.