GPT-6 Astra Triggers OpenAI's Top Cybersecurity Risk Tier, Raising Questions About Untested Rivals
OpenAI's internal safety review placed GPT-6 Astra at its highest cybersecurity risk level, a designation that says more about the gaps in rival model evaluations than about Astra's own capabilities.
OpenAI has assigned GPT-6 Astra its highest internal cybersecurity risk classification, according to reporting from Towards Data Science, marking the first time a company model has landed in that top tier. The designation follows the company's Preparedness Framework, which sorts models by how much damage they could enable in high-stakes domains such as offensive cyber operations. What matters here is less what Astra can do and more what the rating exposes about the rest of the field: most competing models have never been put through an equivalent test.
OpenAI introduced the Preparedness Framework in late 2023 as a way to grade models on four tracked categories, including cybersecurity, before and after deployment. Under that system, models that cross a defined capability threshold face additional safeguards, restricted access, or a halt on release until mitigations are in place. The highest tier is reserved for systems that could plausibly assist a skilled actor in discovering or exploiting serious software vulnerabilities at scale. Astra reaching that level means OpenAI's own evaluators believe the model clears a bar the company has previously treated as a red line.
The disclosure arrives amid a broader shift in how frontier labs handle cyber-capable models. Anthropic publishes capability thresholds for its Claude line, and Google DeepMind has outlined similar tiers for Gemini, but the tests, thresholds, and reporting formats differ across companies. That makes cross-lab comparison nearly impossible, and it means a model from one developer could sit at a dangerous capability level without any public signal. Astra's rating highlights that asymmetry: a top-tier label for one model does little to tell regulators or researchers whether a rival system is safer or simply unmeasured.
Concrete details about Astra remain thin. OpenAI has not published the specific evaluations that triggered the classification, the numeric scores Astra recorded, or which mitigations, if any, the company is applying. The company has previously said it shares summary results with its Safety Advisory Group and, in some cases, with external testers, but it has not committed to releasing the underlying methodology. Without those specifics, independent researchers cannot verify the rating or reproduce it against other models.
Expect pressure to build on two fronts. First, whether OpenAI releases enough technical detail for outside groups to audit the Astra finding, or whether the rating stays a self-reported claim. Second, whether rival labs agree to a common cybersecurity benchmark so that a top-tier label actually means something across the industry. Until then, the headline number tells you where one model sits, not where the field stands.