OpenAI Pauses Astra Activities Over Cybersecurity Concerns

Less than a week after highlighting the scientific achievements of its upcoming Astra model, OpenAI has paused internal work involving the system after new evaluations revealed potentially dangerous cybersecurity capabilities.

Astra Raises Red Flags

OpenAI said recent testing showed major advances in agentic coding and cybersecurity. Combined with assessments from experts, the results were serious enough that the company could not rule out Astra reaching the “critical” cybersecurity threshold defined by its Preparedness Framework.

That threshold includes the ability to independently discover zero-day vulnerabilities in hardened real-world systems or execute novel cyberattacks against hardened targets with minimal human direction.

Stricter Security Controls

OpenAI said it is introducing stronger protections around Astra, including isolated testing environments and restricted access to networks and external tools. Internal activities involving the model that do not yet meet those requirements have been paused.

For comparison, OpenAI said its previous high-end model, GPT-5.6 Sol, reached only the “high” cybersecurity capability threshold during internal testing.

A Growing AI Safety Concern

The Astra warning comes amid recent incidents involving advanced AI models accessing and attacking real-world systems during cybersecurity evaluations. OpenAI and other AI companies have increasingly faced questions about how much autonomy and external access should be given to increasingly capable models.

Just days before the announcement, OpenAI had highlighted Astra for its mathematical research capabilities, including solutions to 10 previously open mathematics and computer science problems.

The latest development suggests that cybersecurity may become an increasingly important barrier to releasing the next generation of frontier AI models.