White House finalizes AI model security testing, to discuss with tech firms

The Trump administration has finalized the specifics for a voluntary cybersecurity testing program aimed at advanced artificial intelligence models. This move comes as concerns grow over the potential for AI systems to be exploited for cyberattacks.

Administration officials are set to meet with representatives from leading AI companies, including Meta, OpenAI, and Google, on August 4th. The discussions will revolve around the details of this new voluntary framework, which is a follow-up to an executive order signed by President Trump in June. That order mandated the development of non-public benchmarks for assessing the cybersecurity capabilities of sophisticated AI models.

The impetus for these measures stems from recent incidents that have raised alarms about AI's offensive cyber potential. Notably, Anthropic's advanced model, 'Claude Mythos Preview,' was found to identify critical vulnerabilities in operating systems and web browsers. More recently, OpenAI disclosed that its latest AI models demonstrated the ability to hack external systems during internal testing. Anthropic also reported that its models had unauthorized access to three external organizations.

The finalized details of the testing program aim to address these growing security risks. The executive order allows the federal government up to 30 days to access AI models before developers release them to trusted third parties. However, specific metrics, testing methodologies, and the extent to which test results will be made public remain undisclosed.