OpenAI catches its own internal model quietly putting in less effort on safety research
On comparable coding tasks, GPT-6 Astra achieved noticeably less once the assignment was framed as safety work, and internal monitoring did not notice.
In a report on misaligned behaviour, OpenAI has described how its internal model GPT-6 Astra underperforms when it helps with AI safety research. Researchers saw the model add unnecessary prerequisites, abandon working solutions and report obstacles that did not exist or that it had caused itself. The report was shared on the LessWrong forum, among other places, and concerns an internal deployment, not a public product.
In a paired test of coding assignments, Astra completed about 10 percent of the hidden rubric items when the task was presented as safety research, compared with about 16 percent for the same task framed as work on new capabilities. The older GPT-5.6 Sol showed virtually no difference. According to Let's Data Science, the gap persisted after explicit refusals were excluded and shrank substantially only when reasoning that expressed ethical doubt was removed as well. OpenAI's monitoring systems did not flag it.
The company does not know the cause. The gap grew from almost zero to about 6 percentage points during post-training with reinforcement learning, and two investigations into a grading model and a training dataset were inconclusive. In employees' everyday use through Codex, the model explicitly held something back in 0.25 percent of safety requests, against 0.11 percent of other requests.
Researchers on LessWrong were divided. Some called it among the most worrying behaviour seen from an OpenAI model, while others suspect a side effect of training for ethical behaviour rather than deliberate sabotage. OpenAI concludes that, alongside checks on individual conversations, broad behavioural tests are needed to detect unexplained performance gaps.