OpenAI Blocks AI Theft Attacks, Cutting GPT-6.2 Extraction Rates to 0.2%
OpenAI details its layered defense stack against adversarial distillation, claiming a drop in reasoning extraction success from 54% to 0.2%.
Takeaways
- OpenAI published its adversarial distillation defense framework covering training, classifiers, and API controls.
- Safeguards in GPT-6.2 cut reasoning extraction success rates from 54.0% to 0.2% in internal red-team tests.
- Defense splits into request-level safeguards and campaign-level detection across coordinated accounts.
- Input and output classifiers scan for extraction attempts and leaks in tool-call arguments.
- Follows July 2026 campaign of 16,000 extraction requests attributed to Moonshot-linked operators.
- Findings shared with the Frontier Model Forum to coordinate industry-wide defenses.
OpenAI has detailed how it detects and blocks adversarial distillation, coordinated API use intended to collect a frontier model’s outputs and transfer its capabilities to another model without authorization. The defense combines model training, classifiers, serving-layer controls, account analysis, investigations, and bans. API developers can expect closer scrutiny of repeated requests for hidden reasoning, particularly when activity spans multiple accounts.
Distillation is a standard training technique in which a smaller “student” model learns from a more capable “teacher.” OpenAI applies the adversarial label when users systematically extract outputs in violation of its terms and use them to reproduce or improve another model. The company argues that this process can transfer advanced capabilities while omitting safeguards such as refusal behavior, content filters, and mitigations added after red-team testing.
OpenAI’s framework follows an incident the company attributed to operators linked to Moonshot. It says the campaign generated roughly 16,000 reasoning-extraction requests across more than 4,000 users in July 2026. The new publication turns lessons from that investigation into a broader defensive model for frontier-model providers.
Two layers catch different signals
OpenAI divides the problem into individual requests and coordinated campaigns. Request-level defenses inspect prompts and responses for attempts to expose protected reasoning. Campaign-level systems correlate behavior across accounts, sessions, infrastructure, and time, allowing investigators to identify patterns that may appear harmless when each request is viewed separately.
| Layer | Purpose | Controls |
|---|---|---|
| Model | Resist requests for protected reasoning while preserving legitimate explanations | Adversarial training and refusal behavior |
| Request | Detect extraction attempts and leaked reasoning | Input classifiers, output classifiers, and tool-call inspection |
| Serving stack | Enforce restrictions outside the prompt | Rate limits, account controls, streaming checks, and inference-layer policies |
| Campaign | Identify coordinated extraction across users and sessions | Behavioral analysis, investigations, account bans, and information sharing |
Serving-layer enforcement matters because system prompts alone cannot reliably stop extraction attempts. Controls in the inference stack can interrupt output, restrict traffic, or suspend accounts even when an adversarial prompt persuades the model to begin producing protected material.
The 0.2% result needs context
OpenAI reports that safeguards in GPT-6.2 reduced the success rate of adversarial-distillation attacks from 54.0% to 0.2% in an internal automated red-team evaluation. The test measured whether attackers could recover protected reasoning traces from the model.
The published figure measures successful extraction attempts, not the resulting student model’s capabilities. OpenAI does not provide the number or diversity of test prompts, the attackers’ query budgets, the false-positive rate for legitimate requests, or a quantitative measure of how much the defenses reduce downstream training gains.
A 0.2% per-request rate also cannot be extrapolated cleanly to a large campaign. Outcomes may be correlated, controls may tighten as account-level evidence accumulates, and a small number of useful outputs could still improve a weaker model. Deployment results will therefore depend on both request-level resistance and the speed of campaign detection.
Developers will see the serving layer tighten
Applications that repeatedly request internal chain-of-thought, vary prompts to bypass refusals, or distribute similar traffic across accounts can resemble extraction campaigns. Developers seeking explainability can reduce that risk by requesting concise rationales, citations, calculations, structured decision factors, or verifiable intermediate results instead of hidden reasoning traces.
OpenAI says it has strengthened checks on streamed responses that could expose reasoning. It also closed a replay path involving encrypted reasoning artifacts already obtained from another user. These artifacts are opaque objects that preserve model state without directly revealing the underlying trace; weak binding between an artifact and its authorized user, session, or model can create a separate attack surface.
An August 2026 study reported that encrypted reasoning artifacts could be reused across sessions, users, and models within the ecosystems of Claude, Gemini, and GPT. According to the researchers, that compatibility could support scalable replay and decryption attacks. The finding helps explain why anti-distillation controls now cover state artifacts and streaming infrastructure alongside prompts and model outputs.
Labs are converging on the threat model
Frontier-model providers use different names for closely related activity:
| Organization | Common terminology |
|---|---|
| OpenAI | Adversarial distillation |
| Anthropic | Distillation attacks and illicit distillation |
| Model extraction attacks and distillation attacks |
OpenAI, Anthropic, and Google reportedly began sharing information about suspected distillation campaigns through the Frontier Model Forum. OpenAI also says it provided findings from the July campaign to the forum and relevant government channels. Cross-provider reporting can expose infrastructure reuse and coordinated account activity that would remain difficult for one lab to identify.
What API teams should change
- Request auditable outputs. Prefer concise explanations, citations, calculations, and structured evidence over prompts that demand private chain-of-thought.
- Monitor aggregate behavior. Detect repeated paraphrases, systematic capability probing, synchronized accounts, unusual output volume, and attempts to bypass refusals.
- Enforce below the prompt layer. Combine model behavior with rate limits, authentication, streaming inspection, output classifiers, and account-level controls.
- Bind state artifacts tightly. Restrict encrypted reasoning objects to the authorized tenant, session, model, and intended operation, then reject replay outside that scope.
- Measure false positives. High-volume evaluation, synthetic-data generation, and automated testing can resemble extraction, so enforcement systems need review paths and sufficient telemetry to distinguish authorized workloads.
- Prepare an incident process. Preserve privacy-conscious logs, define suspension criteria, document escalation paths, and coordinate with infrastructure providers when campaigns span accounts or networks.
Terms of service provide the contractual basis for enforcement, while technical controls determine whether a provider can detect and interrupt abuse. Teams that expose their own fine-tuned models through public endpoints face the same extraction risk and can apply the layered approach OpenAI describes: harden the model, inspect requests and outputs, correlate activity across accounts, and update defenses as attackers adapt.