Skip to content
Theme
OnDevice Tools

AIAI Models & Research

OpenAI makes GPT-6.2 reasoning much harder to extract

According to The Decoder, OpenAI says attacks that extract protected reasoning from GPT-6.2 Astra fell from 54% to 0.2% in its own tests, and it is now banning large numbers of accounts each week.

Illustration of a GPT-6.2 AI model shielded from attackers trying to read its hidden reasoning

Key takeaways

  • OpenAI says attacks that extract protected reasoning from GPT-6.2 Astra fell from 54% to 0.2% in its internal tests.
  • Defences combine targeted training, filters on messages and tool calls, and API blocks that hide reasoning from older models.
  • OpenAI says it suspends a large number of accounts each week, from the thousands into the tens of thousands, to disrupt distillation.
  • Anthropic reports similar action, removing tens of thousands of accounts with a claimed 90% to 99% success rate.
In this article
  1. OpenAI's new safeguards
  2. What changed in the tests
  3. How the defences work
  4. The distillation threat
  5. What it means for developers

OpenAI's new safeguards

OpenAI says it has made GPT-6.2 reasoning extraction much harder for attackers, according to The Decoder. The company reports that attempts to pull the hidden reasoning out of its newest models, GPT-6.2 Astra and GPT-6.2 Sol, fell sharply after fresh safeguards were added. OpenAI is also suspending large numbers of accounts each week to slow coordinated efforts to copy how its models work.

What changed in the tests

Newer OpenAI models produce internal reasoning, a form of step-by-step working that the company treats as protected. Attackers try to recover this reasoning in two ways: by going after earlier requests, or by targeting requests as they are being handled. In its own automated tests, OpenAI says the success rate for the first kind fell from 54% on the older GPT-6 Astra to 0.2% on GPT-6.2 Astra. The second kind fell from 22.9% to 0.9%.

The figures come from adaptive black-box red-teaming, a method in which testers probe a system without seeing how it works inside. OpenAI says it combined results across six attacker models, with lower scores meaning better protection. The company is now applying some of the same improvements to the older GPT-6 Astra.

How the defences work

The safeguards operate on two levels. For individual requests, OpenAI pairs targeted training with filters that inspect both incoming and outgoing messages and tool calls. It also uses API blocks that keep older models from seeing protected reasoning.

For larger, organised efforts, the main defence is account bans. OpenAI says it suspends a large number of accounts each week to disrupt coordinated distillation campaigns. The weekly figure ranges from the thousands into the tens of thousands.

The distillation threat

Distillation is a method where one AI system is trained on the outputs of another. That lets an attacker build a cheaper model that imitates the original. Guarding the internal reasoning is meant to make this kind of copying harder and less faithful.

OpenAI is not the only company facing the problem. According to The Decoder, Anthropic is fighting similar distillation attacks and has also removed tens of thousands of accounts. Anthropic says those efforts succeed in 90% to 99% of cases, and it has urged a joint response from the industry and policymakers.

What it means for developers

Illustration: code, OpenAI logo, AI, Anthropic logo and the labels Anthropic, Distillation, OpenAI

Developers and businesses that build on OpenAI's models are the main audience for these changes. For them, the aim is to protect the work behind the models. Reasoning traces can reveal how a model reaches an answer, and if they leak, rivals may train competing systems at a fraction of the cost.

The reported numbers are OpenAI's own internal results and have not been independently checked. The 0.2% figure should be read as the company's claim, not a confirmed outside measurement. For technical detail and any official guidance, OpenAI's own documentation is the place to look.

Sources

  1. OpenAI makes it much harder for attackers to extract how its latest AI models thinkThe Decoder

Tags

Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.