OpenAI has internally rated its next model — reportedly called Astra — as crossing the company’s “critical” threshold for cyber risk, the first time any of its models has hit that mark. Testing showed Astra can automate cyberattacks, a capability that prompted the company to impose extra security measures before any broader release. The rating, first reported by the Wall Street Journal and confirmed across multiple outlets, signals a meaningful shift in how frontier labs handle models that don’t just answer questions but can take actions in the real world.
The restriction echoes a familiar tension that has dogged OpenAI since the GPT-4 era: the company wants to ship fast, but its own safety frameworks keep throwing up red flags. What’s different here is the specificity. A “critical” cyber rating isn’t a vague concern about bias or misinformation — it’s a concrete, operational risk. A model that can automate attacks at scale changes the economics of cybercrime, and OpenAI seems to recognize that the usual guardrails won’t cut it.
This comes as the Pentagon’s AI apparatus faces its own scrutiny. Emil Michael, the official overseeing military AI, reportedly sold up to $25 million in Perplexity shares — and previously dumped xAI stock after the Pentagon adopted Grok. The optics are lousy, and critics argue he’s running the Defense Department’s AI strategy like a venture capital portfolio. When the people buying AI for national security are also selling AI stock, the line between public interest and private gain gets awfully blurry.
Whether Astra’s restrictions actually hold is another question. OpenAI has a track record of announcing safety measures that later get quietly relaxed under competitive pressure from Anthropic and Google. Google is reportedly shipping Gemini 3.8 Flash imminently, with tests suggesting it could rival OpenAI and Anthropic on coding. If Astra stays locked down while rivals push ahead, the commercial pressure to loosen the reins will be intense. The cyber risk rating might buy time, but restraint has never been the industry’s strong suit.
