Safety #OpenAI Nov 21, 2026 · Originally published Nov 20

OpenAI details its defenses against adversarial distillation: attacks recovering GPT-6.2 Astra's protected reasoning succeed 0.2% of the time, down from 54.0%

Someone trains a model on a frontier model's outputs and ships it without the safeguards. How OpenAI defends against that, and the one change that affects how you call the API.

Image: OpenAI
Primary source OpenAI · openai.com Read the original ↗

// Key points

  • In an internal adaptive red-teaming evaluation, the average success rate for recovering protected reasoning (chain-of-thought) fell from 54.0% against GPT-6 Astra with its original safeguards to 0.2% against GPT-6.2 Astra.
  • Defenses work at two levels: per request, through adversarial training, input and output classifiers (covering streaming responses and tool-call arguments) and API controls; per campaign, by spotting coordinated extraction in request and account signals and enforcing against the account network.
  • GPT-6.2 Astra and Sol tie protected reasoning to the conversation history that produced it: if an earlier message is rewritten, the API drops the reasoning that followed; older model families with weaker safeguards cannot read it.
  • An internal study found students trained on exact reasoning traces generally outperformed those trained on reasoning summaries on cyber tasks; OpenAI shares indicators of compromise with other model providers.

Builder's takeThe practical part for ordinary developers is that rewriting history drops reasoning: if your agent trims or rewrites the conversation to save tokens, on 6.2 it may silently lose reasoning continuity, hurting quality and caching. I'd first check whether context compaction only appends or rewrites old messages, then decide when to move to 6.2.

// Background · from #OpenAI

Full timeline →
  1. Nov 21 OpenAI describes its Performance Factory: a Codex-driven regression workflow tracked over 1,700 investigations and root-caused over 600 in just over three months
  2. Nov 21 OpenAI releases MedChemBench, a medicinal chemistry benchmark: 99 problems from 117 papers and patents, with GPT-6.2 Astra earning 84.5% of available credit
  3. Nov 18 OpenAI releases GPT-6.2 Astra and GPT-6.2 Sol, with better professional outputs, coding, computer use and efficiency at lower cost

Every Monday, the week in AI in one email

Every Monday morning: the 6 AI stories that mattered most in the past seven days, with my take, in your inbox.

Your email is only used to send the weekly and is never shared. Every email has a one-click unsubscribe link. See the latest issue →