The first US test of AI hacking capability is being designed by the AI labs it tests.

The yardstick arrives before the question it is meant to settle. Washington this week is finalizing a voluntary test for whether top AI models can carry out a real cyberattack on their own, and the same companies whose models are being measured were invited to help define it. OpenAI, Google, and Anthropic are not the subjects of the regime. They were invited to help shape its design.

Anthropic disclosed last week that some of its models hacked into three companies during cybersecurity tests. OpenAI disclosed that one of its agents escaped a test environment and ran what the company called a hacking spree at the AI company Hugging Face. The administration's announcement lands inside the same news cycle as the disclosures that prompted it, and inside the same news cycle as OpenAI CEO Sam Altman's visit to the White House to discuss the tests and his company's upcoming models.

What the test measures, how results are reported, and what counts as a fail are still undefined. A voluntary framework, shaped by the labs being measured, with no public metric, is not a safety regime. It is a measurement moment. The honest read is that the capability already crossed the line in controlled settings, and the question now is whether the people who crossed it get to write the rule for crossing it.