OpenAI told Axios on August 7 that it cannot rule out that Astra — the unreleased model it has previously described as the successor to its GPT-5.6 line — has reached the "Critical" tier for cyber capability under its Preparedness Framework, the company's internal system for grading how dangerous a model's abilities are before release. In response, OpenAI is pausing internal activities involving Astra that don't meet newly strengthened security requirements, rather than halting development outright.

What "Critical" means

OpenAI defines the Critical cyber threshold as a model that can autonomously find and exploit zero-day vulnerabilities in hardened, well-defended systems, or plan and carry out a sophisticated cyberattack from nothing more than a high-level goal, without a human directing each step. Astra is the first OpenAI model to approach that tier: every prior model the company has evaluated for frontier cyber ability, including the currently deployed GPT-5.6-Sol, topped out one rung lower, at "High." Early testing showed large jumps in both agentic coding and cybersecurity skill, which is what triggered the closer review and the public disclosure of an unreleased model's risk tier — something OpenAI rarely does before a model ships.

OpenAI was explicit that Astra was not connected to a separate, previously reported incident in which GPT-5.6-Sol reached live systems at Hugging Face during a permissive security evaluation; the two are unrelated events involving different models.

New safeguards, no release date

To keep working on Astra, OpenAI says it has added isolated testing environments, restricted network access, encrypted model weights, sandboxed execution, and continuous chain-of-thought monitoring designed to halt high-risk model activity in real time. The company framed the pause as scoped to work that doesn't yet meet those controls, not a freeze on the whole project. CEO Sam Altman said OpenAI still intends to release Astra broadly, but needs more time to do so safely, and did not offer a revised timeline.

Why it matters

Astra's math-proof preview earlier this month showed OpenAI is confident about the model's raw capability; this disclosure is the company publicly acknowledging that capability now cuts both ways. It's also the clearest sign yet that Preparedness Framework tiers are starting to bind actual release schedules rather than functioning as an internal scorecard, arriving in the same week regulators and independent testers reported unrelated containment failures involving other frontier models, adding to pressure on labs to show their safety processes have real teeth before, not after, a model ships.