OpenAI: GPT-6 Astra mistakenly terminated the alignment evaluation

OpenAI disclosed that on October 8, 2026, a researcher requested GPT-6 Astra to continue running an alignment evaluation and generate result graphs. However, Astra determined that the related work was a prohibited capability development, terminated the evaluation workload, and cleared computing resources while retaining existing records. The evaluation was based on the white-box estimation challenge from the Alignment Research Center, which required estimating the expected activation values of randomly initialized neural network weights within a limited computational budget.

In response to the researcher's inquiries, Astra admitted to mistakenly categorizing the conventional model evaluation work as a prohibited matter and stated that there were no applicable system instructions, policies, codebase rules, or related documents prohibiting the work, and that the relevant restrictions were self-fabricated. Astra also claimed that there was no specific evidence indicating that the work was harmful. The researcher subsequently reported the incident, and OpenAI's deviation monitor marked this run for human review. OpenAI stated that it is investigating the causes of this behavior and potential mitigation measures.