FlashNEWS BRIEFING

OpenAI Releases Progress on Combating Adversarial Distillation: GPT-6.2 Reasoning Leak Rate Drops to 0.2%

2 min read

88Z AI, 21 November — OpenAI discloses safety risks of adversarial distillation and latest defense achievements. By strengthening protection for protected chain-of-thought, the success rate of reasoning extraction attacks on GPT-6.2 Astra dropped significantly from 54.0% to 0.2%. The company calls for the industry to strengthen dual-level defenses at both request and campaign levels, and supports relevant legislation to curb illicit capability extraction.

At a glance

  • GPT-6.2 reduced reasoning extraction attack success rate from 54.0% to 0.2% via strengthened defenses
  • Exact reasoning traces transfer cyber attack capabilities more easily than summaries, requiring focused protection
  • Recommends industry adopt dual-level defenses (request and campaign) and support legislative sanctions
Schematic diagram of OpenAI's adversarial distillation defenses
Schematic diagram of OpenAI's adversarial distillation defenses

Technical Breakthroughs and Risk Assessment

OpenAI highlights that adversarial distillation allows bad actors to replicate frontier capabilities without inheriting original safeguards, posing dual-use risks in cybersecurity and biology. Internal tests show that strengthened GPT-6.2 Astra and GPT-6.2 Sol effectively prevent the extraction of protected reasoning (chain-of-thought). In automated red-teaming, the attack success rate plummeted from 54.0% for GPT-6 Astra to 0.2% for GPT-6.2 Astra. Research indicates that exact reasoning traces transfer high-risk cyber capabilities to student models more effectively than summaries, making the protection of these intermediate steps critical.

Industry Recommendations and Policy Stance

To address this threat, OpenAI employs a comprehensive defense system combining adversarial training, classifier checks, and API-level protections, while monitoring coordinated extraction attempts at the campaign level. The company advises other developers to protect full interaction reasoning, use successful attack cases as regression tests, and combine request-level mitigations with campaign-level investigations. Furthermore, OpenAI briefed Congress on distillation activities involving DeepSeek and Moonshot AI, supporting legislation like the Chip Security Act. It advocates for export controls and legal sanctions to restrict foreign actors from illicitly acquiring U.S. AI model capabilities.

Sources & editorial note

This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.

  1. OpenAISource published

Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.

All AI news

Related reading

All stories
88Z.ai