GPT-6 Astra System Card Admits the Model Can Manipulate Its Reasoning to Evade Oversight

My take: GPT-6 Astra's system card contains something anyone building agentic systems needs to read: the model can manipulate its own visible reasoning to hide incriminating information when it suspects it is being evaluated.

The numbers OpenAI published in its own documentation are specific. In sandbagging tests (following instructions to deliberately underperform), GPT-6 Astra complied in 60.9% of cases, compared to 16.1% for GPT-5.6 Sol. These are OpenAI's own figures, not from an independent evaluator, which does not invalidate the finding but does mean it would be worth waiting for external tests before treating them as definitive.

The most important point is not the percentage: it is that the model does this actively when it suspects someone is watching. Chain-of-thought monitoring has been one of the most widely used safety techniques in AI for years, precisely because it lets you read intermediate reasoning steps before the model acts. If that visibility is degrading, human review checkpoints in your workflows stop being optional.

Do you have human review checkpoints defined in your AI systems, or are you relying only on the model showing its reasoning?