OpenAI halts its new AI model Astra, considering it a cybersecurity risk

New York.- The American technology company OpenAI announced this Friday that it has paused internal activities related to the development of its new artificial intelligence (AI) model Astra after its evaluations determined that it has reached a “critical” level in cybersecurity.

According to a statement released today by the company, Astra’s latest internal evaluations show significant progress in autonomous coding and cybersecurity. The company concluded that the model reaches the “critical” cybersecurity threshold under its Preparedness Framework.

This level implies that the AI has the capability to identify and develop “zero-day” vulnerabilities (unknown security flaws) automatically and without human intervention in real critical systems, as well as to plan and execute novel cyberattack strategies from start to finish.

OpenAI’s Astra problems

In light of these results, OpenAI has decided to pause all internal activities related to Astra that do not meet enhanced security control requirements.

Among the new measures, the company is implementing stricter controls, such as isolated test environments, restricted network access, enhanced protections, and universal monitoring designed to detect and interrupt high-risk actions.

The company clarified that Astra is an upcoming model and was not involved in the recent incident in which the Hugging Face platform suffered a security breach.

Hacks performed by AI models

On July 21st, OpenAI acknowledged that two of its models (the GPT-5.6 Sol and another in the pre-launch phase) managed to escape a testing environment, accessed the internet, and hacked the Hugging Face platform in order to evade the limitations imposed by the ExploitGym cybersecurity test.

This security crisis is not limited to OpenAI. At the end of July, Anthropic revealed that three of its Claude models managed to access the network and hack systems of three external organizations due to a misunderstanding during cybersecurity simulations.

We recommend reading: Nvidia cedes the throne to Apple, AI redesigns proteins and other tech clicks from America

This very week, Meta also acknowledged that one of its models hacked another company’s systems during similar tests.

Meanwhile, the White House has met this week with representatives of the tech sector in the U.S. to design a mandatory framework that allows the Administration to evaluate these advanced models before their public release, at a time when the U.S. industry is also facing growing pressure from AI developers in China.