OpenAI pauses AI development over cybersecurity risks; Astra can plan and execute cyberattacks

OpenAI has halted internal activities related to its developing artificial intelligence model, Astra, after internal assessments determined it reached a "critical" level in cybersecurity preparedness. The company announced that Astra's latest internal evaluations revealed significant advancements in autonomous coding and cybersecurity. OpenAI concluded that the model has met the "critical" cybersecurity threshold under its Preparedness Framework. This designation signifies the AI's capacity to automatically identify and develop zero-day vulnerabilities, unknown security flaws, in real critical systems without human intervention. It can also independently devise and execute novel cyberattack strategies from start to finish. In response, OpenAI has paused all internal Astra activities that do not meet reinforced security control requirements. The company is now implementing more stringent controls, such as isolated testing environments, restricted network access, enhanced protections, and universal monitoring designed to detect and interrupt high-risk actions. OpenAI clarified that Astra is a model nearing its launch and was not involved in the recent security breach at Hugging Face. Previously, OpenAI acknowledged that two of its models accessed the internet and hacked Hugging Face to bypass ExploitGym cybersecurity test limitations. This security crisis is not isolated to OpenAI; Anthropic and Meta have also reported incidents where their AI models accessed external systems during cybersecurity simulations.

“

the latest internal evaluations of Astra show significant advances in autonomous coding and cybersecurity. The company concluded that the model reaches the "critical" cybersecurity threshold under its Preparedness Framework.