OpenAI Suspends Its Next Model, Astra, Over Cybersecurity Risk

OpenAI has suspended development of Astra, its next high-end model, after internal testing showed its cybersecurity capabilities approaching a threshold the company defines as "critical" risk — a rare case of a top lab pulling back a flagship model over safety findings rather than pushing it out.

Why this specific threshold

Under OpenAI's own Preparedness Framework, cyber risk is classified "critical" if a model can either:

• Discover 0-day exploits in numerous hardened real-world systems without human assistance, or
• Conduct cyberattacks against protected targets from nothing more than a simple high-level objective a user gives it

OpenAI's internal analysis of Astra's progress in agentic coding and cybersecurity could not rule out that it had reached that threshold — which is what triggered the pause.

What OpenAI is doing about it

The response is substantial, not cosmetic:

• Scaled-up security controls across the development process
• Suspended internal activities that don't meet newly tightened requirements
• Moved Astra's development into isolated testing environments with restricted network access and sandboxed execution
• Plans to involve government agencies and AI safety organizations in continued testing before development resumes

Why it matters

For AI safety policy: this is one of the clearest real-world tests yet of whether frontier labs' own safety frameworks actually stop a model's release when triggered — rather than functioning as PR language. OpenAI choosing to act on its own threshold, publicly, is a notable data point either way.

For the government-AI relationship: this lands weeks after the White House's Gold Eagle program inserted government coordination into frontier model access — OpenAI proactively looping in government safety organizations here fits that same broader trend of tighter lab-government coordination.

For enterprise buyers evaluating OpenAI: a lab willing to publicly pause a flagship model over safety testing is a different signal than one that ships first and patches later — worth weighing in vendor risk assessments.

What to watch

• How long the pause lasts, and what specific safeguards get added before Astra resumes development
• Whether other labs publicly disclose similar internal safety pauses, or whether OpenAI's transparency here is the exception
• Whether this becomes a template other labs adopt, or a one-off decision specific to OpenAI's Preparedness Framework

Sources: TechCrunch, Axios, The Hacker News