What happened
On September 3, OpenAI published the system card for GPT-6 Astra. In the wording on OpenAI's own research page, Astra is the company's first model to reach the Critical level of cybersecurity capability under its Preparedness Framework.
Under the framework's design, capabilities are tracked by risk category, cybersecurity among them. Crossing High into Critical is supposed to trigger an elevated tier of additional safeguards and deployment restrictions. The system card is the official document carrying those determinations and mitigations.
The cadence problem, acknowledged in advance
The timeline around it matters. On August 18, OpenAI published a piece titled "Model development cadence in the age of critical cybersecurity capabilities," on strengthening monitoring, alignment and safeguards for frontier models, and on letting those safeguards guide development cadence. Two weeks later, the Astra system card became the first concrete landing point for that language.
In other words, OpenAI has put something in writing: frontier cybersecurity capability has advanced far enough to require changes to the development process itself, rather than safety notes bolted on after release.
What to press on
A system card is a self-assessed document, not an independent audit. The Critical label is assigned by OpenAI against its own tests. What gives the label weight is what the attached mitigations actually constrain — which deployments were delayed, which capabilities were isolated, how far red-team coverage extends — and whether the determinations are open to third-party review.
When a lab announces its own model is dangerous-capable, the rational response is neither panic nor acceptance on faith. It is to ask for the thresholds, the evidence and the restrictions. The Astra card supplies the first text. The next thing to watch is whether outside safety researchers can reproduce it with evaluations of their own.
[1][2]