GPT-6 Astra Ships: A Perfect ExploitBench Score and OpenAI's First Model to Trip the Critical Cyber Threshold
What do you do when the lock-picking machine you just built aces its own exam? OpenAI's answer: bolt it in a vault and hand the key only to registered defenders.
Three Key Facts
On September 3, OpenAI officially released GPT-6 Astra — the first model in the company's history to trip the "Critical" cybersecurity capability threshold under its own Preparedness Framework. The official definition is blunt: with the right tools and access, Astra can discover previously unknown security flaws and develop working exploits against well-protected systems without human intervention.
The numbers hit harder than the adjectives. Per The Hacker News, Astra scored 100% on ExploitBench, a benchmark that measures turning known vulnerabilities into functional exploits; the previous generation, GPT-5.6 Sol, managed 78.5%. For high-severity vulnerabilities disclosed between June and August 2026, Astra generated runnable exploit code for 39.0% of them — 3.4x the prior model's 11.5%. During pre-release evaluation it also uncovered two previously unknown zero-day vulnerabilities on its own. On the defensive side, it blocks 91.5% of jailbreak attempts, up from 59%.
Access is tiered: the public model refuses to produce proof-of-concept exploits at launch, while advanced offensive capabilities are restricted to vetted participants of OpenAI's Daybreak program, for defensive work such as vulnerability validation and malware analysis.
WangDou's Take
The most telling thing about this launch isn't the AGI rhetoric — it's the report card itself: an exploit-writing model with a perfect score, protected by a guest list. The locksmith forged a master key to every safe in town, then announced: relax, the key only goes to people who signed the register. The trouble is, 39% of high-severity holes can already be weaponized by machine; what holds back the other 61% isn't a technical wall — the model just hasn't practiced enough yet. And once it has, attack and defense will run on the same brain and the same capability, and the only difference will be who got access first. So don't read Astra as a model upgrade. It's a repricing of the entire offense-defense market: cybersecurity from now on isn't about how many engineers you employ — it's about whether you're on the list. Companies off the list will find the cost of running naked compounding monthly.
Source: The Hacker News