OpenAI Makes It Far Harder For Rivals Like DeepSeek To Copy ChatGPT
OpenAI says it has made it far harder for rival AI labs to secretly copy its models, cutting the success rate of one copying attack from 54% to 0.2%, after earlier linking such attempts to China's DeepSeek and people tie
OpenAI says it has made it far harder for rival AI labs to secretly copy its models, cutting the success rate of one copying attack from 54% to 0.2%, after earlier linking such attempts to China's DeepSeek and people tied to Kimi maker Moonshot AI.
Key Takeaways
- OpenAI said one attack that tries to pull out its models' hidden reasoning now works 0.2% of the time on GPT-6.2 Astra, down from 54.0%
- The company said exact reasoning traces can transfer more cyber capability than reasoning summaries
- Weekly enforcement against distillation campaigns has ranged from thousands to tens of thousands of accounts
- OpenAI linked adversarial distillation to cybersecurity, biology, and national security risks
OpenAI published the research on Nov. 20 and described adversarial distillation as unauthorized use of a frontier model's outputs to train another model with similar capabilities, then deploy it without comparable protections.
The company put special weight on exact reasoning traces, the internal steps a model generates while working toward an answer, because those traces can contain information the final answer leaves out and can become especially valuable training data for another model.
How The Defenses Work
OpenAI described a two-level safeguard stack. At the request level, it uses adversarial training for model behavior, classifier checks, and application programming interface (API) controls to limit reasoning-trace extraction. At the campaign level, it looks for request and account patterns that can reveal coordinated extraction attempts, then investigates and enforces against networks of accounts.
The classifier layer covers both inputs and outputs. Input classifiers look for attempts to extract protected reasoning. Output classifiers look for leaks of protected reasoning in assistant messages and tool-call arguments, and those checks apply to both streaming and non-streaming responses. Tool-call coverage matters because a model can expose protected reasoning in a tool's arguments without putting it in the final answer.
API protections were also tightened across user, workspace, organization, and model-family boundaries. OpenAI said it closed a pathway that could expose another user's encrypted reasoning. With GPT-6.2 Astra and GPT-6.2 Sol, protected reasoning is tied to the conversation history that produced it. If an earlier message is rewritten, the API drops the reasoning that followed it. Older model families with weaker safeguards are also blocked from reading GPT-6.2 Astra and Sol protected reasoning.
In adaptive black-box testing, OpenAI measured attack success rate (ASR) on two kinds of reasoning-extraction attempts and reported lower results on GPT-6.2 Astra in both cases.
| Reasoning Extraction Test | GPT-6 Astra ASR | GPT-6.2 Astra ASR |
|---|---|---|
| Earlier Request | 54.0% | 0.2% |
| Current Request | 22.9% | 0.9% |
The company said the evaluation was aggregated across six attacker models, and that lower ASR is better. It is now applying selected safeguard improvements to GPT-6 Astra as well.
Also Read: OpenAI Locks Down AI Training As Stolen Models Cost Near Zero To Copy
Why This Matters
OpenAI framed distillation as more than an intellectual property dispute. Distillation is a common machine-learning technique in which one model learns from data produced by another, but the company argued that unauthorized distillation can transfer frontier capabilities without transferring the safeguards that govern how those capabilities are used.
That risk is clearest in dual-use fields named in the research. A model that helps defenders find software vulnerabilities could also help an attacker exploit them. A model that assists biological research could also help someone develop a harmful pathogen. OpenAI's internal cyber study found that student models trained on exact reasoning traces generally outperformed students trained on reasoning summaries after supervised fine-tuning. After reinforcement learning, the exact-trace student still scored higher at the largest inference budget tested on ExploitBench and SEC-bench Pro, while the summary-trained student scored higher at lower budgets.
The company also cast adversarial distillation as a national security issue. It said a foreign actor that reproduces advanced capabilities from a United States model can deploy them without the protections its developer put in place. In February, OpenAI said it briefed Congress on activity indicating DeepSeek was attempting to distill United States frontier models. In September, it reported disrupting a campaign to extract protected reasoning and attributed a core part of that activity to individuals associated with Moonshot AI, the developer of Kimi. OpenAI also said it supports the Chip Security Act, the AI OVERWATCH Act, and the MATCH Act, and argued that stronger export controls and distillation defenses both help maintain a United States lead in frontier AI.
What To Watch Next
OpenAI reported seeing multiple unauthorized adversarial distillation campaigns on its platform this year. Weekly enforcement actions have ranged from thousands to tens of thousands of accounts, with the rate increasing significantly this past summer. The company said it tracks both the prevalence of distillation abuse and the time between detection and enforcement.
Its stated goal is to constrain the volume and value of data that distillers can obtain, raise the cost of those campaigns, and shorten the time needed to respond to new attacks. OpenAI also said it plans to share more soon about recent campaigns it has disrupted.
For other model developers and platforms, the company pointed to several steps from this work. Those include protecting private reasoning across the full interaction, checking assistant messages and tool-call arguments for leaks, keeping successful extraction attacks as regression tests, combining request-level mitigations with campaign-level investigation and enforcement, and coordinating safeguards with deployment partners. OpenAI said it has also shared findings through the Frontier Model Forum and appropriate government information-sharing channels.
Read Next: OpenAI Finds 948 Linux Bugs That Could Give Hackers Control
Originally published on Fathom News.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.