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.
More from Models
- Antigravity Student AI Pro users to get same models as basic tier after Nov 30 — Primary-Original2364 · 2026-11-22
- Gemini 4 Carbon vs Neon tested: 48 min/$7.67 vs 34 min/$6.45 on 3D tower build — thebuggeddev · 2026-11-22
- Paying Google Ultra users locked out of Antigravity despite active subscription — AIFlow_ML · 2026-11-22
- T3 Code data shows Neon pulling users away from Opus within days of launch — jeremyphoward · 2026-11-22
- Dev who spent 8+ hours with Neon says no other model comes close to it — rakyll · 2026-11-22
- Is Gemini 4 Neon actually good, or just Theo's YouTube hype? Reddit asks real users — Sorry-Pay-8439 · 2026-11-22