1 min read
Add as a preferred source on Google

OpenAI Says GPT-6.2 Astra Cut Reasoning Extraction Success to 0.2%

Internal automated testing measured attempts to recover protected reasoning traces in GPT-6 Astra and GPT-6.2 Astra.

A sealed metal aperture in a dim testing chamber / TokenPost.ai
A sealed metal aperture in a dim testing chamber / TokenPost.ai

OpenAI’s internal evaluation using automated red-teaming found a 0.2% success rate for attempts to recover protected reasoning traces from GPT-6.2 Astra, compared with 54.0% for GPT-6 Astra with its original safeguards.

The company evaluated the models by testing whether attackers could recover protected reasoning traces. The comparison covered GPT-6 Astra with its original safeguards and GPT-6.2 Astra with updated protections.

OpenAI’s defenses combine adversarial testing, model training, classifiers, API protections and investigations into coordinated extraction attempts. The protections cover GPT-6.2 Astra and GPT-6.2 Sol, though no separate test result for Sol was provided.

Adversarial distillation uses the outputs of a capable model to train or reproduce another model. The practice poses particular risks in dual-use fields such as cybersecurity and biology, where a resulting system could lack the original model’s safeguards.

OpenAI disclosed a separate adversarial distillation campaign on Sept. 30, saying it had disrupted the activity by July 28. The company attributed a core cluster to people associated with Moonshot AI but said it could not establish that every operator belonged to one actor.

Simon Yoon

Reporter

Simon Yoon reports on blockchain technology for TokenPost. Send corrections or tips to info@tokenpost.com.

Loading…