OpenAI: combatting adversarial distillation
Confirmed
The takeaway
On 20 Nov 2026 OpenAI published research on “adversarial distillation”, training rival models on its outputs to copy capabilities without the safeguards.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 22 November 2026
Your AI and this story
- GPT-6.2 Astra173 days after its cutoff
- Claude Opus 5.5143 days after its cutoff
- Gemini 4 Hydrogen234 days after its cutoff
- Grok 4.7173 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 143 days before it.
Key facts
- Red-team results: recovering protected reasoning from an earlier prompt fell from 54.0% to 0.2%, and for the current request from 22.9% to 0.9% (GPT-6 Astra with old safeguards vs GPT-6.2 Astra with new ones)
- Why it matters to OpenAI: in a cyber study, a student model trained on exact reasoning traces generally beat one trained on summaries (ExploitBench, SEC-bench Pro at the largest inference budget)
- Defences: adversarial training, input/output classifiers that also check tool-call arguments, API rules restricting reuse of protected reasoning across users, workspaces, organisations and model families (closing a path that could expose another user’s encrypted reasoning), blocking older model families from reading GPT-6.2 reasoning
- Enforcement: weekly actions ‘from thousands to tens of thousands of accounts’, rising significantly over the summer; indicators shared with other providers
- Named: DeepSeek (Congress briefed in February) and individuals associated with Moonshot AI (September report)
- Policy: supports more US tools to sanction foreign actors that extract capabilities, plus the Chip Security Act, the AI OVERWATCH Act and the MATCH Act
- OpenAI’s X post: about 1.0M views
- Usage Policies effective 20 Nov 2026 (read in the original): ‘If you operate a model-access intermediary platform or other service that enables third parties to directly access OpenAI models, you must obtain our prior review and approval’; apps that merely integrate OpenAI models are exempt. The policy also bars using OpenAI models or outputs to ‘generate, prepare, select, or supply training material or signals’ for alternative models, because training on outputs ‘can transfer capabilities without carrying those safeguards over’ (the training clause dates from the 21 Oct changelog)
What happened
OpenAI described how others try to copy its models by training on their outputs, especially the hidden reasoning, and what it changed to stop that. The post combines red-team numbers, a small study of how much copied reasoning helps a “student” model in cybersecurity tasks, and enforcement figures.
Why it matters
A copy made this way keeps the capabilities but not the safeguards, so for dangerous skills like cyber offence, distillation is a route around the labs’ safety measures. The post also feeds the US debate on export controls and on Chinese open-weight models.
Sources
3 sources from 2 sites. Numbers match the chips in the text.
3 sources: 3 primary
Primary
- OpenAI: Combatting adversarial distillationopenai.com, official
- OpenAI on Xx.com, official
- OpenAI: Usage policies (effective 20 Nov 2026)openai.com, official
Changes
- Usage-policy approval rule for routers and resellers, effective 20 Nov
- Filed (news agent)