Cybersecurity
OpenAI Strengthens GPT-6.2 Astra Against Reasoning Extraction

Key PointsOpenAI reports adaptive extraction attacks on protected reasoning fell from 54.0% to…
View more
- OpenAI reports adaptive extraction attacks on protected reasoning fell from 54.0% to 0.2% with updated safeguards.
- Exact reasoning traces transfer cyber capabilities more effectively than summaries even when the underlying tasks are permitted.
- Defending reasoning requires controls spanning request classifiers tool-call arguments conversation history and account-level campaign enforcement.
OpenAI on November 20, 2026 announced stronger defenses against adversarial distillation, reporting that the average success rate for adaptive attacks extracting protected reasoning from an earlier request fell from 54.0% on GPT-6 Astra with original safeguards to 0.2% on GPT-6.2 Astra with updated safeguards.
For extraction from the current request, the corresponding average fell from 22.9% to 0.9%, OpenAI reports. The adaptive black-box evaluation aggregates results across six attacker models, with automated red-teamers adjusting their approach as the system responds and seeing only what an external user would see. The strengthened protections cover GPT-6.2 Astra and GPT-6.2 Sol, and selected improvements are also being applied to GPT-6 Astra.
Protected reasoning is the intermediate chain-of-thought a model generates while working toward an answer. These traces can contain information omitted from final responses, making them especially valuable to adversarial distillation, which uses a frontier model’s outputs without authorization to train another system. The resulting model may reproduce dual-use capabilities without the safeguards that constrained the original.
What the Cyber-Distillation Study Found
OpenAI’s internal study compared two training runs from the same pretrained student model, the learner in a setup where a teacher model produces the training data. One run used exact teacher reasoning traces; the other used the user-visible baseline of final responses, reasoning summaries and tool interactions. The fine-tuning data consisted primarily of Astra-level teacher trajectories on coding and defensive cybersecurity tasks. Students were evaluated after supervised fine-tuning and again after lightweight reinforcement learning in cybersecurity and coding environments.
After supervised fine-tuning, the exact-trace student generally outperformed the summary-trained student, OpenAI reports. After reinforcement learning, the exact-trace student scored higher at the largest tested inference budget on both reported benchmarks, while the summary-trained student scored higher at lower budgets.
OpenAI interprets the results as evidence that protecting exact traces can limit cyber-capability transfer even when the underlying tasks are permitted under its usage policies, while saying it will continue studying transfer across different student models and training recipes. Across many such interactions, exact traces can teach skills that generalize beyond the teacher’s original tasks.
Request Controls and Campaign Enforcement
At the request level, adversarial training teaches models to recognize extraction attempts and refuse to disclose protected reasoning while preserving legitimate explanations and assistance. Input classifiers look for extraction attempts, while output classifiers inspect assistant messages and tool-call arguments in streaming and non-streaming responses. Tool-call coverage matters because reasoning can appear in a tool’s arguments without appearing in the final answer; output classifiers can block suspected leaks before they reach the caller.
The updated API ties protected reasoning in GPT-6.2 Astra and Sol to the conversation history that produced it. A model can build on earlier reasoning as messages are added, but rewriting an earlier message drops the reasoning that followed it. Older model families with weaker safeguards cannot read the two models’ protected reasoning, restricting reuse across model-family boundaries.
Campaign-level defenses use request and account signals to investigate coordinated extraction, and attacks uncovered during investigations become examples for safeguard training and testing. OpenAI said weekly enforcement actions against unauthorized distillation campaigns in 2026 have ranged from thousands to tens of thousands of accounts, with the rate increasing significantly during summer 2026. The company tracks abuse prevalence and detection-to-enforcement time, aiming to reduce the data attackers obtain, raise campaign costs and shorten response times.
The announcement follows OpenAI’s September 30, 2026 disclosure of a coordinated campaign first observed on July 1, 2026. The earlier report described 16,000 attempted extraction requests from more than 4,000 users on July 24 and 25, 2026, attributed a core cluster to individuals associated with Moonshot AI, and said it was unclear whether all operators belonged to one actor.
Coordination Beyond OpenAI
OpenAI shares indicators of compromise with other model providers and works with deployment partners on safeguard requirements and coordinated enforcement. Its recommendations for other developers include checking assistant messages and tool-call arguments, restricting encrypted-reasoning reuse across conversations and models, retaining successful attacks as regression tests for future safeguards, and pairing request controls with campaign investigations. The company plans to share more about recently disrupted campaigns.
OpenAI has told Congress it supports the Chip Security Act, the AI OVERWATCH Act and the MATCH Act, alongside broader legislation that would give the U.S. government more tools to identify and sanction foreign actors illicitly extracting capabilities from American AI models.
The company argues that export controls and distillation defenses address different parts of that risk: export controls limit access to hardware needed to train frontier models, while distillation defenses make it harder to acquire some of the same capabilities from models already trained.
For model providers, the practical security boundary extends beyond final answers to how reasoning is reused across conversations, exposed through tool calls and monitored across accounts. The study also suggests that permitted coding and defensive-cybersecurity tasks can create a capability-transfer risk, making interaction-wide protection and campaign enforcement parts of the same defense.












