OpenAI pauses parts of Astra development over cybersecurity concerns
OpenAI has paused some work on Astra, an unreleased model, after an internal review found major progress in agentic coding and cybersecurity. The company said Astra reached its “critical cybersecurity threshold,” a designation meaning the model could independently identify and conduct cyberattacks against traditionally well-protected real-world systems.
Under OpenAI’s Preparedness Framework, created in 2023, reaching that threshold requires additional safeguards. The company said preliminary evaluations showed performance strong enough that it could not yet rule out a Critical capability level. OpenAI is continuing to benchmark and assess Astra while applying stricter security controls and suspending internal activities that do not meet the strengthened safeguards.
OpenAI said Astra was not involved in the earlier exploitation of Hugging Face, which involved a different unreleased model during internal testing. The company said it is working with relevant government agencies and selected AI safety organizations to test Astra’s capabilities. It described the disclosure as an effort to be transparent with the public and with safety and security communities about a possible change in model capabilities.
Written by Sakhanda Wire AI from reporting by TechCrunch. Facts, quotes and full context are in the original article below.