OpenAI makes it much harder for attackers to extract how its latest AI models think

Impact 7 The Decoder

Summary

  • OpenAI has significantly tightened security safeguards to prevent attackers from extracting internal reasoning from its GPT-6.2 Astra and GPT-6.2 Sol models.
  • In internal tests, extraction success rates on prior requests dropped from 54.0 percent on GPT-6 Astra to 0.2 percent on GPT-6.2 Astra.
  • Attacks on active requests similarly fell from 22.9 to 0.9 percent using a combination of targeted training, message filters, and API blocks.
  • OpenAI also bans thousands to tens of thousands of accounts weekly to combat coordinated extraction campaigns, mirroring similar efforts by Anthropic.

Why it matters

Protecting internal model reasoning is critical for AI labs to maintain competitive advantages and prevent systematic distillation and theft of proprietary capabilities.

Who should care