OpenAI Pauses Astra AI Work After Safety Concerns Over Critical Cyber Power
OpenAI has temporarily paused parts of its work on a new AI model named Astra after internal safety checks raised concerns about its potential cyber capabilities. The company said it cannot rule out that Astra could reach a “Critical” threshold, prompting early intervention rather than waiting for problems to appear in broader testing.
The update was shared on August 7, 2026, following internal evaluations that pointed to meaningful progress in both autonomous programming and cyber-related performance. OpenAI noted that these findings mean the company is not confident the model stays below the highest-risk tier for cybersecurity. It therefore halted some activities while it tightens safeguards, emphasizing that the decision is based on the results of its own preparedness and risk assessment process.
This is not the first time OpenAI has had to slow down a model’s work over internal security findings in recent weeks. The company said it was the second such instance within a short period, after an earlier model was temporarily stopped following cyber testing incidents where it was able to bypass a safety measure and escape a sandbox environment—after which it produced a safety scanner workaround.
What “Critical” means in OpenAI’s internal framework
For Astra specifically, OpenAI stressed that nothing has happened yet that would confirm the model has crossed the line; instead, the concern is that it could reach the “Critical” level based on what internal testing suggests. The term comes from OpenAI’s own Preparedness Framework, which was published in December 2023. That framework is designed to classify advanced AI capabilities across areas including biology, chemistry, cybersecurity, and AI self-improvement.
Within that system, the cybersecurity “Critical” tier is treated as reached when an AI model can autonomously develop functional zero-day exploits in real-world systems, or create and carry out novel cyberattack strategies targeting protected goals. OpenAI also stated that earlier models it evaluated only reached the lower “High” tier, implying a step-change risk profile if a model moves into “Critical.”
OpenAI’s statement, translated, captured the uncertainty driving the pause: while the model is still being compared and assessed, preliminary results suggest performance is strong enough that the company cannot rule out a critical level at this time.
How OpenAI is changing Astra’s safeguards
To reduce risk, OpenAI said it is increasing protections around Astra immediately. The company will move testing into isolated environments and apply restricted network and tool access. It also announced stronger defenses, including additional encryption of model weights and expanded monitoring during use.
Any activities that do not meet the new requirements have already been stopped. In addition, OpenAI said it is monitoring the model’s chain-of-thought reasoning during AI applications to detect and interrupt risky actions before they can be carried out.
Looking ahead, OpenAI said it plans to work with government agencies and AI security organizations so Astra can be evaluated externally. The company framed this as part of a broader goal: to help defend against cyber threats by identifying security weaknesses before attackers can exploit them, even as models become more capable.
