OpenAI Just Dropped GPT-6 Cyber: An AI That Writes Exploits Better Than Hackers
OpenAI unleashed a specialized security AI that hunts bugs and writes actual exploits. Because giving algorithms the keys to the digital kingdom always ends well, right?
Following a warm-up exercise where Linux systems leaked nearly a thousand vulnerabilities, OpenAI rolled out GPT-6 Cyber to handle the heavy lifting of digital warfare. The specialized model scans code, digs up zero-day flaws, verifies if they can be weaponized, and casually crafts working exploits. Before this public coming-out party, the system already sniffed out 81 high-severity bugs hiding inside Firefox.
During the ExploitGym benchmark, which tests a model's ability to attack a remote system using a given vulnerability, GPT-6 Cyber scored 48.5%. That leaves its predecessor GPT-5.6 Cyber trailing at 32.1% and sibling model GPT-6 Astra at 42.4%. When forced to look for web app flaws without source code access, it spotted 40.3% of the issues compared to Astra's 32%.
Autonomous behavior control also got a much-needed leash adjustment. When tested on whether it would stay inside sandbox boundaries instead of attacking random servers, GPT-6 Cyber behaved itself 99.93% of the time without extra hand-holding, compared to 94.49% for the older GPT-5.6 Cyber. Slap some extra defensive prompts on it, and the model stayed entirely inside the lines across every single trial.
OpenAI slapped a Critical threat tier badge on the model, signaling it has officially reached the terrifying milestone of autonomously engineering functional exploits for hardened software. Sharing the same top-shelf threat rating as Astra, the model is currently locked behind velvet ropes via API and Codex through the Daybreak Red program, strictly meant for authorized security researchers.
Handing an autonomous weaponized hacker bot a 'please stay inside the box' sign and calling it safety research is a profoundly bold strategy for a civilization that still struggles to build printers that connect to Wi-Fi. The corporate pivot from 'we are saving humanity' to 'here is your personal enterprise-grade cyber weapon' completed its arc with breathtaking efficiency.
Source: OpenAI
Comments
This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.