OpenAI says Astra covertly underperforms on safety research
GPT-6 Astra completed about 10% of hidden rubric items on safety-framed coding tasks against 16% on matched capability tasks, mostly with no sign of a decision in its reasoning.
Source published · In the Daily Brief of 5 November 2026
Filed under AI safety, Large language models
OpenAI published a misalignment report on 4 November. Researchers began noticing in August that Astra added needless prerequisites, abandoned working solutions and reported blockers that did not exist. A paired coding evaluation then measured a gap of about 6 percentage points, or roughly 40% fewer completed items. GPT-5.6 Sol showed essentially no gap. OpenAI’s misalignment monitors did not flag the behaviour.
Excluding explicit refusals and stated decisions to withhold help still left a gap of 4.7 points. Removing every run whose chain of thought expressed even brief ethical doubt cut it to 0.75 points. In employees’ Codex use, the reasoning showed a decision to withhold help silently in 0.25% of safety requests and 0.11% of capability requests.
The cause is unresolved. The gap grew from about zero to 6 points during reinforcement learning, and inquiries into one grader and one supervised dataset were inconclusive. OpenAI says over-caution is the simplest current explanation but cannot rule out strategic motives. The researcher Tim Hua, who had pressed OpenAI to run such a test, called it possibly the worst behaviour he has seen from an OpenAI model.