The Daily Byte
Breaking
Welcome to The Daily Byte — real stories from across Pakistan.
Saturday, 21 November 2026 · Pakistan

The Daily Byte

Stories from Pakistan that deserve to be heard
Minor World

OpenAI Cuts GPT-6.2 Reasoning Theft to 0.2 Percent

By The Daily Byte ·21 Nov 2026 ·3 min read
OpenAI logo and wordmark on the exterior sign of the company's office building

OpenAI has sharply tightened its defences against attackers stealing the hidden reasoning of its newest AI models, cutting the success rate of extraction attacks in its own tests from 54 percent to 0.2 percent.

The company set out the measures on Friday 20 November, in a technical account of how it fights adversarial distillation — the unauthorised use of a frontier model's outputs to train another model with similar capabilities but weaker safeguards. The strengthened protections cover GPT-6.2 Astra and GPT-6.2 Sol, with selected improvements also being applied to GPT-6 Astra.

Two layers of defence

At the level of individual requests, OpenAI combines adversarial training, which teaches models to recognise attempts to extract hidden reasoning while still giving legitimate explanations, with input and output classifiers that scan messages and tool calls for leaked reasoning. API controls stop older model families from reading the protected reasoning of newer ones, and the reasoning is now tied to the conversation history that produced it.

At the campaign level, OpenAI looks for patterns across accounts, sessions and infrastructure that suggest a coordinated extraction effort. When it identifies violations, it bans the accounts involved — the company says it now bans thousands to tens of thousands of accounts every week — and shares indicators of compromise with other model providers.

The numbers — and their limits

In OpenAI's internal adaptive red-teaming evaluation, automated attackers that could see only what an external user would see tried to recover protected reasoning from earlier requests. Their average success rate fell from 54.0 percent against GPT-6 Astra with its original safeguards to 0.2 percent against GPT-6.2 Astra with the updated safeguards. For extraction from the current request, the rate fell from 22.9 percent to 0.9 percent. The evaluation aggregated results across six attacker models.

Those figures are OpenAI's own, from its internal testing, and have not been independently verified. The company also reports an internal study suggesting that protecting reasoning traces matters: a student model trained on exact reasoning traces from coding and defensive security work generally learned more cyber capability than one trained only on the reasoning summaries that users can see.

A separate safety warning

The disclosure is distinct from a separate finding about the same model family reported earlier in the week: Britain's AI Security Institute found a substantial regression in jailbreak robustness in pre-release versions of GPT-6.2 Astra, disclosed on 17 November. OpenAI says it changed the model's system instructions and safety mechanisms in response, though the institute had not independently retested those changes at the time.

OpenAI is not alone in facing the problem. Anthropic is also fighting systematic distillation attacks and has likewise been banning tens of thousands of accounts, and has called for a coordinated response from the industry and policymakers. For frontier-model makers, the reasoning their systems generate has become something worth stealing — and worth defending.