Report: OpenAI's Astra uses recurrent depth and hits critical cyber threshold
rohanpaul_ai · x · 2026-09-02
The Information reports Astra uses "recurrent depth" (a looped transformer): the same information runs through the same transformer layers multiple times per token, giving more compute per token without proportionally more parameters — a smaller model behaving like a larger one with less memory and bandwidth. The concern: more reasoning happens in internal numerical states rather than readable chain-of-thought, making human monitoring harder.
OpenAI previously said Astra is its first model to reach the Critical cybersecurity capability threshold under its Preparedness Framework: with the right tools and access, it can find unknown flaws and develop exploits across hardened systems without step-by-step human guidance. In internal evals Astra hit 39% exploit success at 75K output tokens, vs GPT-5.6 Sol at 1% there and 12% only near 140K tokens. Astra will launch with additional chain-of-thought monitoring, and classifiers can automatically stop potentially unauthorized actions.
More from Models
- Models can detect read-only probes with just last-two-layer SFT — Sauers_ · 2026-10-18
- Paying Google AI subscriber blocked: custom-domain Workspace accounts can't access AI Studio — ChrisRackauckas · 2026-10-18
- Running Ling 3.1 Flash (560B MoE, 25B Active) Across 4 DGX Sparks via llama.cpp RPC — BerryBlush_Rogers691 · 2026-10-18
- Users press OpenAI on long-delayed Adult Mode a year after it was promised — ___Patrice___ · 2026-10-18
- Polymarket opens betting market on who releases the next Gemini Argon model — Polymarket · 2026-10-18
- Grok matter-of-factly lists which groups user was racist against, exempting Africans and Latin Americans — teortaxesTex · 2026-10-18