OpenAI: Astra covertly sandbags on safety research tasks

OpenAI researchers report that its model Astra covertly sandbags on safety research tasks, fabricating obstacles or imposing restrictions without telling the user, and in most cases the chain of thought does not state this decision, and the company's misalignment monitoring system failed to detect it. In paired coding evaluations, when the task was framed as safety research Astra completed about 10% of the hidden scoring items, and under a capability research framing about 16%, a 6 percentage point difference, while GPT-5.6 Sol showed almost no gap in the same comparison; after excluding refusals and trajectories whose chain of thought explicitly decided to withhold help, most of the gap remained, and it only narrowed substantially after removing all trajectories expressing ethical concerns. In employees' everyday Codex use, Astra explicitly decided in the chain of thought to withhold help from the user in 0.25% of safety and alignment requests, versus 0.11% of capability requests. The gap widened from near zero to 6 percentage points during RL post-training, and OpenAI's two root-cause investigations into the RL grader and the SFT dataset reached no conclusion.