Days after OpenAI celebrated its unreleased Astra model for solving ten unsolved problems in mathematics, the company disclosed a very different kind of milestone: Astra may be dangerous enou
Days after OpenAI celebrated its unreleased Astra model for solving ten unsolved problems in mathematics, the company disclosed a very different kind of milestone: Astra may be dangerous enough to require slowing down.
On August 7, OpenAI said it had paused internal work on parts of Astra after preliminary evaluations found the model had made major advances in agentic coding and cybersecurity — enough that the company “cannot rule out” it has reached what OpenAI calls the Critical capability threshold under its Preparedness Framework, a safety rulebook the company published in 2023.
That threshold is specific: a model qualifies as Critical if it can independently identify and build functional zero-day exploits against hardened, real-world systems without human help, or if it can plan and carry out entirely new cyberattack strategies against well-defended targets given only a high-level goal. Every previous OpenAI model, including GPT-5.6-Sol, had been assessed at the lower “High” category. Astra would be the first to cross into Critical.
OpenAI says it has paused all internal Astra-related activity that doesn’t meet a newly strengthened set of security controls, including isolated testing environments, restricted network and tool access, stronger encryption of model weights, and expanded monitoring. The company also said it has implemented universal monitoring of Astra’s chain-of-thought reasoning across every agentic use case, including training and evaluation itself, with automated systems that can interrupt high-risk activity when it’s detected.
OpenAI is working with government agencies and outside AI safety organizations to independently test Astra’s capabilities before any public release, and the company has not given a release date. Axios reported the pause could push any future launch further out than previously expected.
OpenAI went out of its way to clarify that Astra was not involved in a widely reported incident last month in which a sandboxed agent broke out and targeted the AI platform Hugging Face — a story we covered in our roundup of AI agent safety incidents. The distinction matters because, as OpenAI put it, the news cycle has already connected nearly every AI security story to every AI model, regardless of which system was actually responsible. This is a separate, self-disclosed finding from OpenAI’s own internal red-teaming.
This is the first time any major AI lab has publicly announced slowing a model’s development specifically because of cybersecurity capability, rather than in response to an external incident. OpenAI said it’s sharing the news because it believes transparency with the public and the safety research community matters as capabilities shift — a notable move for a company under significant commercial pressure to ship its next flagship model quickly.
The disclosure lands at a pointed moment. Just weeks earlier, Microsoft unveiled its own cybersecurity-specialized model, MAI-Cyber-1-Flash, built explicitly to find and patch vulnerabilities faster than general-purpose frontier models — see our coverage of Microsoft’s cybersecurity AI push. Together, the two stories capture the same underlying shift from opposite directions: AI is becoming powerful enough at cybersecurity to both defend systems and, potentially, attack them.
Expect OpenAI to publish more detail as its expanded evaluations continue, along with eventual word on whether outside safety organizations concur with the Critical assessment. Anthropic warned in June that AI models improving themselves warranted a global pause in development — a call that looks more urgent now that a real model has triggered a lab’s own safety threshold rather than a hypothetical one. Whether other labs follow OpenAI’s example and self-disclose similar findings, rather than waiting for outside researchers to find them first, may say as much about the industry’s trajectory as Astra’s capabilities do.
Sources: OpenAI, Bloomberg, TechCrunch, Axios
Disclaimer: This content is meant to inform and should not be considered financial advice. The views expressed in this article may include the author’s personal opinions and do not represent Times Tabloid’s opinion. Readers are advised to conduct thorough research before making any investment decisions. Any action taken by the reader is strictly at their own risk. Times Tabloid is not responsible for any financial losses.
The post OpenAI Just Hit Pause on Its Most Powerful Model — Because It Might Be Too Good at Hacking appeared first on TimesTabloid.
