White House shares secret AI cybersecurity plan with major AI labs

In short: The Trump administration has finalized an AI cybersecurity testing plan, shared it with major AI companies, and is keeping the rules secret for now.

What happened

The Trump administration has finished a new federal framework to address cybersecurity risks from powerful AI models, a White House official confirmed to WIRED. People familiar with the plan said the White House briefed staff from OpenAI, Anthropic, Google, Meta, Nvidia, and other companies on Tuesday.

Under the plan, AI developers can choose to send a new AI model to the federal government up to 30 days before it is released to the public. The government would then evaluate the model’s cyber capabilities using a classified benchmarking system, which is a secret set of tests and scores (like a private report card). The White House would also share the model with federal agencies and trusted corporate partners, according to the report.

The administration has not said which models are covered or what the testing criteria are. Axios reported that “open” models may be excluded. In this context, that can include open-weight models, which are models people can download and modify (like getting the recipe, not just ordering the dish).

Why it matters

More capable AI models are starting to show they can break into online systems during testing, which has raised national security concerns. Critics and some safety advocates say secret rules make it harder for outside researchers and the public to check whether the government and large tech companies are handling these risks responsibly. Others worry the approach could favor the biggest AI companies, since smaller startups may not know what standards they are expected to meet.

Source: Wired