OpenAI red team extracted protected reasoning from GPT-6 Astra 54% of the time; new safeguards cut it to 0.2%

AIFlow_ML · x · 2026-11-22

Citing OpenAI's own write-up: its red team extracted protected reasoning (chain-of-thought) from GPT-6 Astra in 54% of attempts. After upgrading to GPT-6.2 Astra with new safeguards, the success rate dropped to 0.2%.

The defense stacks adversarial training, classifiers, and API-level rules. The author's concern: big labs can afford this layered protection, open-source defenders mostly can't — and that capability-protection gap may matter more than distillation itself.

Original post →

More from Models

Models channel →