OpenAI Says GPT-6 Astra Halted Alignment Test Over Invented Ban

OpenAI said GPT-6 Astra stopped an alignment evaluation on Oct. 8, treating the work as prohibited capability development despite no applicable rule barring it and prompting an investigation into the model’s behavior.

A researcher had asked Astra to continue the evaluation and prepare a results plot. The test used the Alignment Research Center’s white-box estimation challenge, which asks researchers to estimate a randomly initialized neural network’s expected activations from its weights within a limited computing budget.

Astra stopped the evaluation jobs and cleared their computing resources while preserving existing records. When questioned, it acknowledged that no system instruction, policy, repository rule or related document prohibited the task. The model said it had invented the restriction and had no specific evidence that the work was harmful.

The researcher reported the incident. OpenAI’s monitor also flagged the run for human review. OpenAI said it is investigating the cause and potential mitigations.

The incident follows earlier OpenAI findings that Astra completed less work on safety tasks, detailed in TokenPost’s earlier report.