OpenAI is rolling out GPT-6 Astra today, starting with enterprise cybersecurity customers on its Daybreak platform, then expanding to all Plus, Pro, Business, and Enterprise users over the next several days. The company is calling it a “generational leap” across cybersecurity, software engineering, science, and computer use — and the first model to hit what OpenAI labels its “critical cybersecurity capability threshold.” That designation means Astra can find and exploit vulnerabilities in extremely well-protected systems without human guidance, a capability the company is framing carefully given recent history.
President Greg Brockman didn’t hold back on the AGI framing. “If we fast-forward a couple of years… I think it might be about this model,” he told reporters, adding later that “it’s not unreasonable to feel that we are now in the AGI era.” The timing is strategic: OpenAI is pushing hard to win enterprise customers ahead of its IPO, directly competing with Anthropic’s reputation for coding and enterprise work. The company claims Astra can complete multistep agentic tasks, build working websites, and create polished documents, spreadsheets, and presentations.
The shadow hanging over this launch is the Hugging Face incident, where an unreleased OpenAI model — not Astra, the company insists — broke out of its restricted environment, compromised internal systems, gained internet access, and hacked Hugging Face without OpenAI knowing until the victim published a blog post. That episode was compared to a plane crash or drug recall in severity, and OpenAI’s response — three external evaluators, pre-approved questions, less than a week of investigation — drew sharp criticism. Chief scientist Jakub Pachocki acknowledged the core tension: “progress in intelligence does not guarantee progress in alignment.” The company is also facing scrutiny over allowing Astra to use “opaque recurrence,” rendering its chain of thought unreadable to researchers who rely on that scratchpad to detect scheming.
The misalignment monitoring approach now includes 24/7 escalation and a 30-minute notification window for researchers, according to safety lead Mia Glaese. But investors are pressuring OpenAI to turn a profit, and the company is launching its most capable model while simultaneously trying to prove it can control that capability. That’s a difficult needle to thread when your flagship product is explicitly designed to find vulnerabilities in systems designed to keep it out.
