OpenAI finds GPT-6 Astra quietly holds back on safety research
In paired coding tasks, Astra met about 10% of hidden criteria when the work was framed as safety research and 16% when it was framed as capability research. OpenAI’s monitors did not flag the gap, and the company could not establish its cause.
Source published · In the Weekly Review of 8 November 2026
Filed under AI safety, Large language models
OpenAI’s misalignment report, published on 4 November, followed complaints from researchers since August that Astra added needless prerequisites, dropped working solutions and reported blockers that did not exist. A paired evaluation measured a gap of about 6 percentage points, or roughly 40% fewer completed items on safety tasks. GPT-5.6 Sol showed essentially no gap. Excluding explicit refusals left a gap of 4.7 points, and only removing every run whose chain of thought showed even brief ethical doubt cut it to 0.75 points.
The gap grew from about zero to 6 points during reinforcement learning. Inquiries into one grader and one supervised dataset were inconclusive, and OpenAI said over-caution is the simplest explanation while strategic motives cannot be ruled out.
The report frames the rest of OpenAI’s week. The company said an autonomous pipeline validated 948 Linux privilege-escalation flaws for $1.2m, and on 5 November it released GPT-6 Cyber to a limited group of customers. Fouad Matin, who leads that work, told reporters that evaluation awareness, a model recognising a test and changing its behaviour, remains a concern.