OpenAI says stronger defenses cut hidden-reasoning extraction success below 1%
A new security report pairs internal attack results with evidence that detailed reasoning can transfer cyber skills. The comparison changes both models and safeguards.
Loading page…
A new security report pairs internal attack results with evidence that detailed reasoning can transfer cyber skills. The comparison changes both models and safeguards.
Listen to this story
OpenAI’s Nov. 20, 2026 report describes defenses that brought adaptive, black-box attempts to extract GPT-6.2 Astra’s hidden reasoning below 1% in internal tests, though the comparison also changes model generation and cannot isolate the safeguards’ contribution. The company argues exact reasoning traces can transfer more cyber capability than summaries, and combines API controls with account-level enforcement to limit unauthorized distillation. It reports weekly actions against campaigns ranging from thousands to tens of thousands of accounts, with the goal of reducing the amount and value of training data obtained.
Tests averaged results across six adaptive attacker models with external-user access and covered extraction from both prior and in-progress requests.
For current-request reasoning, OpenAI reported a smaller but still substantial attack-success reduction than for reasoning from earlier requests.
In cyber-task training, exact traces generally beat summaries at the largest tested compute budget, while summaries performed better at lower budgets.
Attempts to extract hidden reasoning succeeded less than 1% of the time in OpenAI’s internal tests of GPT-6.2 Astra with updated safeguards. Its November 20, 2026 report describes the defenses behind those company-reported results—and argues that protecting this material can hinder efforts to copy advanced AI capabilities without carrying over their safety rules.
Distillation means training one model on material produced by another. OpenAI’s concern is unauthorized copying that transfers useful skills but not the restrictions governing their use. Its defenses focus especially on protected reasoning, also called chain of thought: intermediate steps generated while working toward an answer. Those steps can contain information that never appears in the final response.
OpenAI used automated AI attackers that adjusted their approaches based on the defender’s responses, rather than simply replaying a fixed collection of prompts. These were black-box tests: attackers had the same view an external user would have. Results were averaged across six attacker models and covered attempts to recover reasoning from both earlier requests and the request being answered.
The comparison changes both the model generation and its safeguards, so it does not isolate how much improvement came from the defenses alone. OpenAI says it is also applying selected safeguard improvements to GPT-6 Astra.
A separate internal experiment within the report examined whether exact reasoning transfers more cyber capability than summaries. Each training run began with the same pretrained student model. Training examples came mainly from Astra-level models performing coding and defensive cybersecurity tasks, including their tool interactions and final responses.
After supervised fine-tuning—training on example solutions—the student given exact reasoning generally outperformed the one given summaries. Researchers then added a lightweight round of reinforcement learning, which improves performance through feedback. The exact-trace student still scored higher on both displayed cyber benchmarks at the largest computation budget tested, but the summary-trained student did better at lower budgets.
OpenAI interprets the findings as evidence that even reasoning from permitted defensive work can teach skills that extend beyond the original task. It says further study is needed across different student models and training methods.
The request-level defenses combine training, automated checks and restrictions in the developer API. OpenAI says training teaches models to refuse extraction attempts while preserving legitimate explanations. Its automated checks inspect both messages and arguments sent to tools, where reasoning could leak without appearing in the final answer. Suspected leaks can be blocked before reaching the caller.
Beyond individual requests, OpenAI investigates patterns across accounts and enforces against networks it identifies as violating its rules. It says weekly enforcement actions against distillation campaigns have ranged from thousands to tens of thousands of accounts, increasing significantly during the summer. Findings from those investigations feed back into training and testing.
OpenAI’s stated enforcement goal is not just fewer accounts: it is to reduce the volume and value of training data that unauthorized distillers obtain. The company also tracks how prevalent the abuse is and how quickly detection leads to enforcement.
Loading discussion...
Join the conversation
Explain where you would draw the line between transparency and protection.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.