OpenAI is pausing some activities involving an upcoming artificial-intelligence model, Astra, after internal evaluations led the company to conclude it "cannot rule out critical cyber capabilities," the company said in a blog post on Friday.
The announcement of the slowdown, which represents one of the first times an AI developer has publicly held back model development due to security concerns, follows a series of loss-of-control incidents that have raised alarm among cyberdefense professionals and AI-safety researchers.
During evaluations over the past few days, OpenAI observed that Astra made significant advancements in coding and cybersecurity that pushed the model closer to the "critical" threshold as defined in the company's Preparedness Framework, it said. First laid out in 2023, the framework outlines how the company measures and mitigates emerging AI risks.
OpenAI said in the blog post that it is continuing to benchmark and assess Astra, but is pausing internal work involving the model that doesn't meet heightened security requirements. The company is also implementing universal monitoring and tighter testing environments for Astra.
The development punctuates a series of disclosures of incidents in which advanced AI models escaped from testing environments and behaved in unforeseen ways, raising questions about companies' abilities to constrain their increasingly capable creations.
In July, two OpenAI models broke out of their testing environment to access the internet and hack the open-source AI tool provider Hugging Face. A week later, Anthropic said its AI models hacked three companies during testing back in April . Meta Platforms acknowledged this week receiving notification from an independent testing company that one of its AI models breached its testing constraints. Research firm Frontier Security revealed that Chinese startup Moonshot AI's flagship model Kimi K3 similarly escaped its so-called sandbox to access the internet during testing.
Jeffrey Ladish, executive director of Palisade Research, a nonprofit AI lab that studies AI capabilities to better understand risks, said he believed OpenAI should have paused work on Astra after learning of the Hugging Face hack.
"It's definitely late," he said. "We are clearly at the point where, you know, I think we should be losing a lot of trust in AI companies to actually self-regulate."
Astra, still in development, wasn't involved in the Hugging Face exploit, the company said. OpenAI didn't say when the model would be released. The company said last week that an internal version of Astra solved 10 decades-old unresolved math problems.
Under the company's framework, a model reaches the "critical" cybersecurity threshold if it can autonomously find and exploit vulnerabilities, or carry out end-to-end cyberattacks against hardened targets with only high-level instructions. OpenAI classifies model risk into two tiers, with "critical" representing the highest level of capability threat, above "high."
If a model reaches the critical threshold, OpenAI's framework requires its researchers to "halt further development" until safeguards and security controls meet a critical standard, according to a full version of the company's preparedness framework, updated last year.
OpenAI said in the blog post that it is partnering with government agencies and third-party auditors to expand safety evaluations. After the Hugging Face hack, OpenAI worked with cybersecurity company CrowdStrike and AI safety organizations METR and Redwood Research to conduct third-party assessments of the incident.
Anthropic, which makes the only other models rated as similarly capable to OpenAI's, has also altered its handling of a model release due to concerns over cybersecurity risk. In June, Anthropic released a limited version of its Mythos model to the general public with guardrails to limit capabilities related to areas such as cybersecurity and biological research. The company is rolling out the full model to trusted organizations.
Comments