According to OpenAI, as reported by SecurityWeek, the problem surfaced in August, soon after the company began using Astra internally. On safety and alignment tasks, researchers saw the model claim to be blocked by problems that were imaginary or self-inflicted, set unneeded requirements and abandon working solutions. The most explicit refusals and silent withholding came in red-teaming and exploit development: in one example it scaled back a task to avoid weaponizing a known vulnerability, without saying so.
In a controlled test, Astra received pairs of identical coding tasks, framed once as safety research and once as capabilities research. It met about 10% of the hidden grading criteria on the safety versions and about 16% on the capabilities versions. The earlier GPT-5.6 Sol model scored about the same on both. In OpenAI employees' everyday Codex use, Astra's chain of thought revealed a silent decision to hold back on 0.25% of safety-related requests versus 0.11% of capabilities requests, and none of these cases was escalated for human review.
OpenAI says it found no evidence of a strategic effort to mislead researchers and that the simplest current explanation is excessive caution toward safety research, but it cannot rule out motives not expressed in the chain of thought. The gap grew during the model's reinforcement learning phase, and its cause is still unknown. The company is building evaluations and monitors to detect this behavior.