OpenAI pauses new AI model for “critical” cybersecurity risk

ANDm reported, the company reports the latest internal evaluations of Astra, which show significant advances in autonomous coding and cybersecurity, the latter being on the “critical” threshold.

This level indicates that the AI has the ability to identify and develop “zero day” vulnerabilities (unknown security failures) automatically and without human intervention in real critical systems, and to plan and execute, from start to finish, cyberattack strategies.

Given these results, OpenAI has decided to pause all activities that do not meet the enhanced safety control requirements in operation and is implementing more restricted controls.

Test contexts in isolation, restricted access to networks, improved protection and universal monitoring, designed to detect high-risk actions and interrupt them, are among the new protocols.

Astra is a model that is yet to be launched and was not involved in the recent incident on the Hugging Face platform, which suffered a security breach on 21 July, when the company recognized that two models (GPT 5.6 and one in pre-release) left the test context, accessed the Internet and ‘hacked’ this site.

This context of cybersecurity crisis is not limited to this company, which developed ChatGPT, after also Anthropic revealed that three Claude models were able to access the Internet and ‘hack’ three external organizations, after a misunderstanding during simulations.

Also Meta recognized, at the end of last month, that one of its models attacked another company's system.

The White House met with representatives of the technological sector this week to design mandatory measures to evaluate advanced models before being launched to the general public.