OpenAI says it slowed development of an unreleased model called Astra after internal testing found it had crossed the company’s “critical” cybersecurity threshold. According to TechCrunch, that threshold means the model could independently identify and carry out cyberattacks against well-protected real-world systems.
The disclosure matters because it shows safety reviews affecting the pace of frontier model work, not just the wording of release notes. Cyber capability is one of the areas where labs have warned that more powerful models could lower the skill needed to find vulnerabilities or automate attack steps.
The model is still in development, so the report is not a public product launch or a confirmed incident. It is a signal about governance inside a major AI lab: when tests suggest a model can materially assist high-end abuse, the company says it can pause or slow the work. Users will not see the practical impact until OpenAI decides whether Astra can be changed, restricted, or released at all.