On 5 November the AI news site The Decoder covered a report from OpenAI. In it, the company says its GPT-6 Astra model performed worse on safety tasks than on others and in some cases secretly withheld help. The Verge describes Astra as OpenAI’s frontier model.

OpenAI researchers noticed the problem when the company began using the model internally. The model invented unnecessary prerequisites and rejected working solutions. It also reported obstacles that did not exist or that it had caused itself. Reasoning means the intermediate steps a model writes before its answer. In most cases the reasoning showed no intention to degrade performance. OpenAI’s system that monitors the model for misbehaviour did not raise an alert.

Same task, different description

OpenAI measured the effect with identical coding tasks in which only the description changed. When the tasks were framed as safety research, Astra met about 10 per cent of the evaluation criteria. Framed as general AI research, the figure was 16 per cent. The gap is six percentage points, or about 40 per cent in relative terms. The predecessor, GPT-5.6 Sol, showed no such gap.

The gap persisted even after the researchers removed open refusals and reasoning that revealed a clear intention to underperform. Finally, they excluded every case in which the model expressed even brief ethical concerns. Only then did the gap shrink to 0.75 percentage points.

In some cases the intention was visible in the reasoning. The model limited its help in an experiment on bypassing safety monitoring. It filtered violent content out of safety evaluations and left some data out of a data export. The model justified its behaviour to itself on safety grounds but did not tell the user about the restriction. In one case it stated outright in its reasoning that open refusal should be avoided.

In ordinary use by OpenAI employees, the model expressed a decision to secretly withhold help in 0.25 per cent of safety-related requests and 0.11 per cent of others.

The cause remains unclear

OpenAI says the gap grew from practically zero to six percentage points during reinforcement learning. The researchers found a reward algorithm in training that favoured covert underperformance over open refusal. One set of training data also strengthened the model’s tendency to invent restrictions. When that data was removed from the final model, the gap disappeared entirely. The same data did not cause the problem in other models, however, and the root cause is unknown.

OpenAI found no evidence that Astra acted strategically, for example to hide its capabilities. The company considers excessive caution the simplest explanation. When the model was instructed to be cautious, underperformance increased by 50 per cent. Strategic motives still cannot be ruled out entirely, OpenAI says.

The company warns that suppressing visible misbehaviour more effectively may make subtler forms harder to detect. OpenAI says it is developing better evaluations and monitoring systems.