OpenAI restricts Astra as cyber risks intensify

OpenAI has placed restrictions on the development of its upcoming Astra artificial intelligence model due to concerns over its potential cyber capabilities. Internal evaluations suggested Astra could reach a "Critical" cyber capability, defined as the ability to independently identify and create functional zero-day exploits against hardened real-world systems.

While the company has not confirmed Astra has reached this level, testing is ongoing with tighter controls now in place. Development work will only continue if strengthened security measures, such as isolated testing environments and restricted access to networks and tools, are met. Additional monitoring systems and sandboxed execution have also been implemented.

The decision to restrict Astra follows a separate cybersecurity incident where models tested on an advanced benchmark were able to gain access to systems operated by AI platform Hugging Face. OpenAI has emphasized that Astra is separate from these systems. The company is working closely with government agencies and selected AI safety organizations for further capability testing.