Skip to content

OpenAI Details Defenses Against Adversarial Distillation

2 reports 2 reporting sources first reported Updated 9 hours ago

Understand the event

Report summary

OpenAI details its defenses against attempts to extract protected reasoning and use it to train models without comparable safeguards. In an internal adaptive black-box evaluation aggregated across six attacker models, the average success rate for extracting reasoning from an earlier request fell from 54.0% on GPT-6 Astra with original safeguards to 0.2% on GPT-6.2 Astra with updated safeguards. OpenAI also describes request-level protections, campaign investigations, and enforcement.

From OpenAI News

Read next

More changes in today’s briefingHugging Face and Thinking Machines Lab Release Inkling 2 Mini

Report timeline

Follow the reports to explore different sides of the event.

11/21
  1. OpenAI Shares Research on Detecting and Disrupting Model Distillation

    OpenAI shares research on efforts to combat distillation of its models, examining how capabilities transfer and how the company detects and disrupts coordinated attempts. It says the findings are informing its defenses and supporting safeguards across the industry.

11/20
  1. How OpenAI Is Combating Adversarial Distillation

    OpenAI details its defenses against attempts to extract protected reasoning and use it to train models without comparable safeguards. In an internal adaptive black-box evaluation aggregated across six attacker models, the average success rate for extracting reasoning from an earlier request fell from 54.0% on GPT-6 Astra with original safeguards to 0.2% on GPT-6.2 Astra with updated safeguards. OpenAI also describes request-level protections, campaign investigations, and enforcement.

Keep up with AI, one email a day

Model releases, price changes and compute developments, with original sources. Free, confirmed by email, and easy to unsubscribe. Read a past issue

Event heat trend

Current within comparable coverage
16
Peak within comparable coverage
1911/21 05:00 (UTC+8)
24-hour change
–

Trends compare the same continuously observed participants, which may cover fewer sources than current heat. Point or click to inspect hourly heat, or use left and right arrow keys.