OpenAI Cuts Reasoning-Extraction Success From 54% to 0.2% in GPT-6.2 Astra
OpenAI says updated safeguards cut adaptive reasoning-extraction success from 54.0% to 0.2% on GPT-6.2 Astra, as weekly enforcement hits tens of thousands of accounts.
By Daniel Okafor
4 min read
What's News
- Extraction success on earlier requests fell from 54.0% (GPT-6 Astra) to 0.2% (GPT-6.2 Astra), OpenAI reported November 20, 2026.
- Current-request extraction fell from 22.9% to 0.9% across a six-attacker-model black-box evaluation.
- Weekly enforcement actions in 2026 ranged from thousands to tens of thousands of accounts, accelerating in summer 2026.
- OpenAI attributed a core cluster of a 16,000-request July 2026 campaign to individuals associated with Moonshot AI.
- Exact reasoning traces transferred cyber capabilities better than summaries even on permitted tasks.
By Daniel Okafor
Adaptive attacks extracting protected reasoning from an earlier request fell from 54.0% to 0.2%, OpenAI reported on November 20, 2026. The drop compares GPT-6 Astra under original safeguards with GPT-6.2 Astra under updated defenses. For extraction from the current request, the average success rate fell from 22.9% to 0.9%, the company said.
The evaluation was a black-box test. Six attacker models probed the system, automated red-teamers adjusted their approach as the system responded, and the attackers saw only what an external user would see. The strengthened protections cover GPT-6.2 Astra and GPT-6.2 Sol, and selected improvements now extend to GPT-6 Astra.
Why reasoning traces matter more than answers
Protected reasoning is the intermediate chain-of-thought a model generates on the way to an answer. These traces can contain information omitted from final responses. That makes them valuable to adversarial distillation, the practice of using 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.
OpenAI's internal study tested what exact traces actually teach. It compared two training runs from the same pretrained student model. One run used exact teacher reasoning traces. The other used the user-visible baseline: final responses, reasoning summaries and tool interactions. The fine-tuning data consisted primarily of Astra-level teacher trajectories on coding and defensive cybersecurity tasks.
After supervised fine-tuning, the exact-trace student generally outperformed the summary-trained student. After lightweight reinforcement learning, the exact-trace student scored higher at the largest tested inference budget on both reported benchmarks. The summary-trained student scored higher at lower budgets.
OpenAI reads the results as evidence that protecting exact traces can limit cyber-capability transfer even when the underlying tasks are permitted under its usage policies. Across many interactions, exact traces can teach skills that generalize beyond the teacher's original tasks. The company said it will continue studying transfer across different student models and training recipes.
What do the new controls actually do?
At the request level, adversarial training teaches models to recognize extraction attempts and refuse to disclose protected reasoning while preserving legitimate explanations. Input classifiers look for extraction attempts. 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, which restricts reuse across model-family boundaries.
Campaign-level defenses combine request and account signals to investigate coordinated extraction. Attacks uncovered in investigations become examples for safeguard training and testing. OpenAI said weekly enforcement actions against unauthorized distillation campaigns in 2026 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.
A coordinated campaign preceded the fix
The announcement follows OpenAI's September 30, 2026 disclosure of a coordinated campaign first observed on July 1, 2026. That report described 16,000 attempted extraction requests from more than 4,000 users on July 24 and 25, 2026. It attributed a core cluster to individuals associated with Moonshot AI, and said it was unclear whether all operators belonged to one actor.
How the defense extends 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:
- Check assistant messages and tool-call arguments
- Restrict encrypted-reasoning reuse across conversations and models
- Retain successful attacks as regression tests for future safeguards
- Pair 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 giving the U.S. government more tools to identify and sanction foreign actors illicitly extracting capabilities from American AI models. It argues export controls and distillation defenses address different parts of the risk: export controls limit access to hardware, while distillation defenses make it harder to acquire capabilities from models already trained.
For model providers, the security boundary now extends beyond final answers to how reasoning is reused across conversations, exposed through tool calls and monitored across accounts. The study suggests permitted coding and defensive-cybersecurity tasks still create capability-transfer risk, which makes interaction-wide protection and campaign enforcement parts of the same defense.
Source: unite.ai
More from Daniel Okafor
Show full bio
Correspondent covering money, funding and market shifts around entrepreneurship, startups and venture funding at Business Bearings.
2,538 articles