Pick

OpenAI reduces success rate for extracting earlier reasoning from 54.0% to 0.2% in GPT-6.2 Astra

·2026.11.20 19:00

Key point

Updated protections in GPT-6.2 Astra and Sol reduce the success rate for extracting protected reasoning from earlier requests from 54.0% to 0.2%.

1 / 3

Details

  • OpenAI reports that updated safeguards in GPT-6.2 Astra and GPT-6.2 Sol reduced the adaptive black-box attack success rate for extracting protected reasoning from earlier requests from 54.0% to 0.2%.
  • Internal studies show that training student models on exact reasoning traces transfers more cyber capabilities than training on reasoning summaries, particularly after reinforcement learning at high inference budgets.
  • Defenses combine adversarial training, input and output classifiers, and API-level controls that tie protected reasoning to specific conversation histories and prevent cross-model access.
  • OpenAI has disrupted multiple unauthorized distillation campaigns, with weekly enforcement actions ranging from thousands to tens of thousands of accounts, and supports legislation to sanction foreign actors extracting U.S. model capabilities.

Technical Defenses

  • Adversarial training uses successful attacks from red-teaming and investigations to teach models to refuse reasoning extraction while maintaining legitimate responses.
  • Output classifiers detect leaks in assistant messages and tool-call arguments, blocking suspected protected reasoning before it reaches the caller.
  • API protections restrict reasoning reuse across conversations and models, dropping reasoning if earlier messages are rewritten and preventing older model families from accessing GPT-6.2 reasoning.

Industry and Policy Implications

  • OpenAI shares indicators of compromise with other model providers and supports a consistent-standards policy for coordinated enforcement across deployments.
  • The company supports the Chip Security Act, AI OVERWATCH Act, and MATCH Act to strengthen export controls and limit adversary access to training hardware.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.