SIAgentNews.com
← All news

OpenAI cuts reasoning extraction attacks on GPT-6.2 Astra to 0.2%

New safeguards aim to prevent competitors and bad actors from siphoning off the internal decision-making processes that make frontier models valuable.

· First reported by Crypto Briefing

0.2%Success rate of reasoning extraction attacks

The short version

  • OpenAI's new safeguards reduced reasoning extraction success from 54.0% to 0.2% for earlier conversation requests.
  • The company uses adversarial training and API-level locking to protect hidden 'chains of thought.'
  • OpenAI is sharing indicators of compromise with other model providers to fight distillation as a shared security threat.
  • Lower extraction risks may increase confidence in using AI for sensitive fields like cybersecurity and biology.

Why it matters

By making it harder to copy the 'how' behind an AI's answer, OpenAI protects its competitive advantage and intellectual property. It also moves the industry toward a collaborative security model where labs share data on bad actors.

The threat of adversarial distillation

OpenAI is currently fighting a process known as adversarial distillation. In the AI world, distillation refers to training a smaller, cheaper model using the outputs of a larger, more capable one. While this is a standard practice when done with permission, it becomes a security risk when done without authorization.

The specific target of these attacks is protected reasoning. This is the internal, step-by-step thinking a model performs before it produces a final answer. OpenAI keeps this chain of thought hidden from users because it contains the most valuable information about how the model functions.

If an attacker can successfully coax this hidden reasoning out of a model, they gain a training signal that is much richer than a simple final answer. As reported by Crypto Briefing, this allows outsiders to essentially copy the internal logic of a frontier model at an industrial scale.

Measuring the drop in attack success

OpenAI recently published research on November 20, 2026, to measure how well their new safeguards work. The study compared the older GPT-6 Astra model against the newer GPT-6.2 Astra. The results showed a dramatic decrease in the success rate of attacks designed to pull protected reasoning from the system.

For attacks targeting reasoning tied to earlier requests in a conversation, the success rate plummeted from 54.0% on the old model to just 0.2% on the new one. For reasoning tied to the current request, the success rate dropped from 22.9% to 0.9%. These figures represent a significant hurdle for anyone trying to siphon off the model's internal logic.

While a 0.2% success rate does not completely eliminate the possibility of extraction, it fundamentally changes the economics of the attack. It makes each successful extraction much more expensive and difficult for a determined actor to achieve at scale.

How the new safeguards work

The company is using two primary methods to defend against these extraction attempts. The first method is adversarial training. This involves deliberately exposing the model to extraction attempts during its initial training phase so it can learn to recognize and resist them.

The second defense is implemented at the API level. OpenAI now ties a model's reasoning to specific conversation histories. This means that the reasoning generated in one specific exchange is locked to that exchange only. An attacker cannot easily pick up that reasoning and replay it in a different context to fish for more information.

OpenAI has also taken active enforcement measures throughout 2026. The company reported taking weekly actions against thousands of accounts that were identified as being involved in distillation campaigns.

Broader security and industry impact

These defenses are not just about protecting OpenAI's business interests. The research suggests that lower adversarial risk could increase trust in using these systems for sensitive work. In fields like biological research or cybersecurity, leaking a model's internal reasoning could provide dangerous capabilities to the wrong people.

OpenAI is also treating distillation as a shared security problem rather than a private fight. The company has begun sharing indicators of compromise with other model providers. This allows rival labs to spot the same actors attempting to exploit their own platforms.

By distributing this information, OpenAI aims to foster a more collaborative environment. This approach helps strengthen security across the entire AI sector by identifying and blocking bad actors who move from one lab to another.

The road ahead for model security

The research highlights that while the improvements are significant, some areas remain more vulnerable than others. The success rate for current-request reasoning, which sits at 0.9%, is still higher than the 0.2% rate for earlier-request reasoning. This suggests that attackers may shift their focus to these more accessible points of pressure.

OpenAI's recent release of GPT-6.2 Astra and GPT-6.2 Sol coincided with these security updates and a 20% price cut for Astra's API. These moves signal a push to make high-quality, secure AI more accessible while simultaneously hardening the walls against unauthorized copying.

As the industry matures, the battle between model creators and those trying to distill their intelligence will likely intensify. The success of these new safeguards will determine how well companies can protect their intellectual property while still providing powerful tools to the public.

What to watch

  • Whether the 0.9% success rate for current-request reasoning becomes a primary target for new distillation techniques.
  • How other model providers respond to OpenAI's initiative of sharing indicators of compromise.

Try it

Check the latest OpenAI API documentation to see how they structure conversation histories and reasoning locks.

Sources and links

Read the original at Crypto Briefing

First reported by Crypto Briefing. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news