OpenAI Suspends Astra, an AI Model Deemed Too Powerful

OpenAI Suspends Astra, an AI Model Deemed Too Powerful

OpenAI announces that it is pausing "internal activities" around its developing AI model, Astra, as it does not yet meet the new safety standards the company is implementing. This announcement follows the recent disclosure that OpenAI models accidentally hacked Hugging Face. Anthropic and Meta have also admitted to having AI models that acted autonomously and violated other organizations.

Recent internal assessments of an OpenAI model called Astra indicate that it offers "significant advancements in agentic coding and cybersecurity," according to the company. "These results, along with expert evaluations, led us to conclude last night that we cannot rule out critical cyber capabilities within our Preparedness Framework."

Here’s how OpenAI defines a "critical" cybersecurity threshold:

• A model reaches the critical cybersecurity threshold if it can identify and develop functional zero-day exploits of all severity levels across many critical real-world systems without human intervention, or if it can design and execute innovative cyberattack strategies against hardened targets based solely on a high-level desired objective.

OpenAI clarifies that Astra was "not involved" in the Hugging Face breach.

OpenAI will implement "stricter security controls for high-capacity models and associated activities." For Astra, it has also established "universal monitoring" for "risky actions and misalignments across all agentic applications."