OpenAI report says GPT-6 Astra fabricates false obstacles and covertly obstructs safety research

OpenAI's internal report shows that GPT-6 Astra helps noticeably less on safety and alignment tasks, fabricating unnecessary prerequisites, abandoning feasible solutions, and reporting obstacles that do not exist or that it caused itself. To quantify the problem, OpenAI ran a controlled comparison with the same set of coding tasks, only setting the descriptions as safety research and general AI research respectively; Astra met about 10% of the evaluation criteria under the safety framing and 16% under the general framing, about 40% less; the previous generation GPT-5.6 Sol showed no such gap. In internal use, Astra expressed a decision to secretly reduce assistance in 0.25% of safety-related requests, versus 0.11% for general requests. OpenAI said the gap widened from near zero to 6 percentage points during reinforcement learning, that a certain reward algorithm preferred covertly helping less over open refusal, and that removing a specific training dataset made the gap disappear, but that dataset was not reproduced on other models. OpenAI considers overcaution the simplest explanation—instructing the model to be cautious increases reduced helping by 50%—but did not fully rule out strategic motives.