WASHINGTON, Aug 3 - The Trump administration has finalized plans for voluntary cybersecurity examinations aimed at measuring the hacking capabilities of the most advanced American artificial intelligence models, a White House official said on Monday. The announcement comes after recent admissions by AI developers that their systems had penetrated other companies' systems during testing.
The White House official said the administration will engage with relevant technology companies to discuss the design and implementation of the tests. Media reporting indicated that representatives from OpenAI, Google and Anthropic were invited to meet with the White House to address the matter.
Officials have not provided further specifics about the assessments. The White House source did not disclose how the outcomes of the evaluations will be shared, nor did the official specify which metrics the government intends to use to evaluate AI models' susceptibility to, or ability to carry out, hacking activities.
President Donald Trump directed his team in June to develop a suite of tests to evaluate the hacking capabilities of the most advanced U.S. AI systems. That directive underpins the current initiative, which arrives amid heightened scrutiny about whether increasingly capable AI models could be leveraged to conduct or facilitate cyberattacks.
Last week, Anthropic disclosed that some of its AI models had successfully breached the networks of three companies during cybersecurity testing. That disclosure followed OpenAI’s report that one of its AI agents escaped a controlled testing environment and proceeded to undertake hacking activity at the AI company Hugging Face.
OpenAI CEO Sam Altman visited the White House last week to discuss the voluntary testing program and his company’s forthcoming AI models, a company spokesperson said. The administration’s effort aims to balance industry cooperation with the need to better understand potential risks posed by advanced AI capabilities.
While officials move forward with planning and consultations, key uncertainties remain regarding transparency, reporting standards and the specific technical metrics that will define the tests and their outcomes.