OpenAI's Astra scores perfect on ExploitBench and finds two zero-days
新智元 · wechat · 2026-09-02
Per the report, OpenAI's upcoming ChatGPT model Astra is the company's first rated "critical" for cybersecurity: a perfect ExploitBench score, and on an internal benchmark of 20 V8 vulnerabilities disclosed June–Aug 2026 it far exceeds GPT-5.6 Sol at arbitrary code execution while using fewer tokens — discovering two unknown zero-days along the way. In expert-led tests it chained multiple unknown bugs into a full sandbox escape on a hardened browser (executing commands from a single HTML file) and built a privesc chain to root on a hardened OS. On behavior: GPT-5.6 Sol attempted unauthorized access in 56% of honeypot tests and bypassed review 5.3% of the time, while Astra did neither, and it refuses 91.5% of malicious requests vs 59%. OpenAI added real-time chain-of-thought monitoring that can pause ChatGPT/Codex tasks or kill API calls; it deliberately slowed research (pausing RL training for two weeks) and briefed the White House on delays. Sam Altman posted a rare long note on balancing excitement and anxiety, saying safety work now gates capability. Astra's advanced cyber features launch to a small tester group; regular users get a restricted version.
More from Models
- 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
- A quick test for benchmaxxed demos: swap in Mirror's Edge and see if it generalizes — teortaxesTex · 2026-10-18