OpenAI Sells “Hacking” AI to Security Firms — Days After Pausing a More Powerful One

In a move that has raised eyebrows across the tech world, OpenAI has launched GPT-5.6-Cyber — its first “offense-grade” artificial intelligence model built specifically to hunt for security flaws, validate exploits, and conduct advanced vulnerability research.
The launch comes just three days after the company paused its next-generation model, Astra, over fears it was approaching a “critical” level of hacking capability under OpenAI’s own safety framework. GPT-5.6-Cyber, by contrast, sits one notch below that threshold and is being sold — under strict vetting — to established cybersecurity firms including CrowdStrike, Cloudflare, Accenture and IBM.
The results are striking: on OpenAI’s internal benchmark, the new model completed 95% of complex exploit-chain tasks, compared to just 1.5% for the standard version of GPT-5.6. It has already uncovered previously unknown security flaws in Google Chrome’s browser engine and in a widely used mobile operating system.
The launch is part of OpenAI’s expanded “Daybreak” program, which now offers two tiers of access for vetted cybersecurity defenders — a sign that AI labs are increasingly treating powerful offensive capabilities as a controlled product rather than simply refusing to build them at all.

OpenAI Sells “Hacking” AI to Security Firms — Days After Pausing a More Powerful One