Editorial illustration for OpenAI cuts AI model attack success from 22.9% to 0.9%, ports fixes to GPT-6 Astra
OpenAI Slashes AI Attack Success Rate to 0.9%
Extracting an AI model's internal reasoning used to be fairly easy. OpenAI's own numbers show how far that's changed with GPT-6.2 Astra and GPT-6.2 Sol, the company's latest models. Attacks that once pulled protected reasoning from prior requests succeeded 54 percent of the time against GPT-6 Astra. Against the new version, that figure dropped to 0.2 percent.
The stakes here are less about curiosity and more about theft. Rival labs and researchers have long tried to extract chain-of-thought output from frontier models, both to understand how they work and, more damagingly, to train competing systems on stolen reasoning traces, a practice known as distillation. OpenAI isn't alone in treating this as an active threat. Anthropic has spent months banning accounts tied to similar extraction campaigns, and has pushed for industry-wide coordination on the problem.
What's notable about OpenAI's latest move is the combination of defenses at play, covering single requests and larger, coordinated attack campaigns alike. The company is also retrofitting some of those protections into GPT-6 Astra rather than reserving them for new releases. The internal test results, run through adaptive red-teaming, give a sense of how much the attack surface has actually shrunk.
In internal automated tests, attacks pulling protected reasoning from prior requests dropped from a 54.0 percent success rate on GPT-6 Astra to 0.2 percent on GPT-6.2 Astra. Attacks on active requests fell from 22.9 to 0.9 percent, OpenAI says.
Why this matters
For anyone building on top of GPT-6 Astra or GPT-6.2 Sol, this is worth tracking closely. A drop from 54.0 percent to 0.2 percent on stored reasoning extraction, and from 22.9 to 0.9 percent on live requests, tells us OpenAI's prompt-injection and reasoning-leak problem was bigger than most developers assumed. These aren't edge-case exploits; they're attacks that apparently worked on more than one in five active sessions before the fix. That's the kind of number that should make founders building agentic tools on these models go back and audit what they've already shipped.
The decision to backport fixes to GPT-6 Astra rather than treat GPT-6.2 as a clean break is the more interesting signal. It suggests OpenAI sees this as a platform-wide liability, not a one-model quirk. For researchers, the combination of targeted training, input/output filters, and API-level blocks on older models is a useful template for how model providers are starting to patch reasoning leakage the way we patch memory vulnerabilities. We'd still want independent red-team numbers before taking OpenAI's internal 0.9 percent at face value.
Common Questions Answered
How much did OpenAI reduce the success rate of attacks extracting protected reasoning from prior requests?
OpenAI reduced the attack success rate from 54.0 percent on GPT-6 Astra to 0.2 percent on GPT-6.2 Astra, representing a dramatic 99.6 percent reduction in vulnerability. This improvement demonstrates significant progress in protecting the model's internal chain-of-thought reasoning from extraction attempts.
What is the difference between attacks on prior requests versus active requests in GPT-6.2 Astra?
Attacks pulling protected reasoning from prior requests dropped to 0.2 percent success rate, while attacks on active requests fell to 0.9 percent success rate. The distinction reflects two different attack vectors: one targeting stored reasoning from previous interactions and another targeting live ongoing sessions.
Why is the reduction in prompt-injection and reasoning-leak vulnerabilities significant for developers building on GPT-6 Astra?
The previous 22.9 percent success rate on active requests revealed that reasoning-leak attacks worked on more than one in five sessions, indicating a widespread vulnerability rather than an edge-case exploit. This substantial reduction to 0.9 percent provides developers with significantly improved security when building applications on top of these models.
How does GPT-6.2 Sol benefit from the security fixes implemented in GPT-6.2 Astra?
OpenAI ported the reasoning-extraction attack fixes from GPT-6.2 Astra to GPT-6.2 Sol, ensuring both latest models receive the same security improvements. This approach allows developers using either model to benefit from the reduced vulnerability to chain-of-thought extraction attempts.
Further Reading
- Combatting adversarial distillation - OpenAI
- GPT-6.2 System Card - Deployment Safety Hub - OpenAI
- OpenAI cuts reasoning-extraction success to 0.2% on GPT-6.2 Astra - The Circuitry
- OpenAI Strengthens GPT-6.2 Astra Against Reasoning Extraction - Unite.AI
- OpenAI Says New Defenses Cut Reasoning Theft From GPT-6.2 Astra to 0.2% - AI Parabellum