Industrial safes with manual locks and keys

OpenAI Says Its New Reasoning Locks Cut Extraction Attacks From 54% to 0.2%

OpenAI Says Its New Reasoning Locks Cut Extraction Attacks From 54% to 0.2%

THE HOOK

OpenAI has bolted the vault door on its models' chain-of-thought, and it has numbers to prove it.

THE GIST

In its write-up on combatting adversarial distillation, OpenAI says average extraction attack success against protected reasoning fell from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra. Defences include classifiers, adversarial training and API limits binding reasoning to its conversation.

THE TAKE

Reasoning traces are gold for anyone training a copycat, so locking them down is rational. It's also convenient for a company that sells those tokens. The figure comes from OpenAI's own red-team, so treat it as a claim, not a verdict. Attackers will adapt.

OPINION

This is an arms race wearing a safety badge: the stated worry is distilled cyber capabilities, the commercial worry is rivals cloning the margin. Silicon Report has a useful summary, but independent testing is what would convince me.

  • Story type: Important story
  • Impact: Medium
  • Confidence: Medium (company-reported results)
  • Audience: Developers
  • What changed: OpenAI hardened protected-reasoning defences in GPT-6.2 Astra and Sol.
  • What to watch: Independent red-team results and whether rivals publish comparable defences.
  • Reading time: 1 minute