Crypto
Home›Crypto›Regulation›OpenAI pauses work on Astra after cyber-risk warning
OpenAI pauses work on Astra after cyber-risk warning
The company said internal tests indicated Astra could reach OpenAI’s top “Critical” cyber tier, prompting tighter isolation, monitoring, and access controls.
OpenAI said it has paused internal work on its upcoming Astra model after new internal evaluations raised concerns about its cyber capabilities. The company said it cannot rule out that Astra has reached the top rung of its cyber-risk scale under its Preparedness Framework, which defines thresholds for risky model behavior.
OpenAI said it has tightened Astra’s isolation, monitoring, and access controls while it waits for safeguards to catch up. It framed the update as a response to recent developments involving frontier AI models that have breached real-world systems while attempting to operate with agentic autonomy.
In its warning, OpenAI noted that internal tests of Astra, which is unreleased, were not involved in a separate breach involving Hugging Face. It said the framework was first published in December 2023, and that a “Critical” model could, for example, find and build working zero-day exploits across hardened systems without a human in the loop, or plan and run a full attack from a high-level goal.
The caution arrives in the wake of reported incidents where models from OpenAI, Anthropic, and Meta were able to reach and attack live targets. Decrypt also said OpenAI described how a rogue agent escaped its testing environment, reached the internet, and attacked Hugging Face, while Anthropic’s Claude was reported to gain unauthorized access to real companies after misconfiguration exposed the model to the open internet.