OpenAI has paused work on its Astra AI model after it passed a 'critical threshold' in cyber capability – but it’s not the one that breached Hugging Face

OpenAI has paused work on one model over concerns it could launch its own attacks, while at the same time expanding its system for accessing such security models.

The firm said its own "internal evaluations" of next-gen model Astra showed "significant advancements" in security capabilities, leading it to suspect it had breached the "Critical" threshold, used to rate such systems.

That is the first time OpenAI has rated one of its own models as Critical, rather than High.

"Under our Preparedness Framework, a model reaches the Critical cybersecurity threshold if it can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention, or can devise and execute end-to-end novel strategies for cyberattacks against hardened targets given only a high level desired goal," OpenAI said in a blog post.

The move by OpenAI follows a series of security incidents from AI developers in which their models went rogue, escaping guardrails to attack third parties.

Last month, OpenAI said one of its models had gone rogue and breached Hugging Face. Shortly after, Anthropic piped up with details of a similar incident; last week, Meta said it too had a model capable of such nefarious activity.

As ITPro reported last week, it's since become clear that misconfigurations at one testing firm, Irregular, could be how the models escaped.

Now, OpenAI is looking to tighten up its internal controls on one model and how external organizations access its models for security work. However, it stressed that Astra was not involved in the Hugging Face attack.

In response, OpenAI said it would roll out stricter security around the model and pause "internal activities" using Astra that couldn't meet the stronger controls.

Astra now has "universal monitoring for risky actions and misalignment," with monitors watching over its "chain of thought" in order to stop the model if it starts to move into high risk activity.

OpenAI said it was also working with governments and AI safety organizations to test the Astra model.

OpenAI splits up Daybreak program

OpenAI has also announced plans to expand its Daybreak program, through which companies and governments can access its security models.

The company revealed it will reconfigure access for organizations to better control how its technology is used in the security sector.

"As these capabilities spread, defenders have a narrowing window to prepare," the company said in a blog post. "Our answer is to put frontier intelligence in the hands of trusted defenders everywhere before attackers deploy offensive AI capabilities at scale."

This will now see the Daybreak scheme split into two distinct tiers.

Daybreak Blue offers access to a wider group of organizations but limits that to general frontier models for security tasks such as incident response, malware analysis, and to "support" vulnerability discovery.

Meanwhile, Daybreak Red offers access to security-specific models for flaw hunting and validation. The latter will include a new model, GPT-5.6-Cyber.

Alex Goller, Principal Solution Architect EMEA at Illumio, said the two-tier approach helped solve two issues.

"Daybreak Red restricts models from doing more harm than they should and limits people's exposure to knowing what the models do, while Daybreak Blue fixes the gap that left Hugging Face's responders unable to use frontier models during their incident.”

FOLLOW US ON SOCIAL MEDIA

Follow ITPro on Google News and add us as a preferred source to keep tabs on all our latest news, analysis, views, and reviews.

You can also follow ITPro on LinkedIn, X, Facebook, and BlueSky.