Read it as a dependency with no contract. Congress has not passed comprehensive AI legislation, and the federal agencies in this file do not agree on where oversight starts or how it should work [5]. That leaves the pre-release arrangements resting on good will, plus whatever levers the executive already holds elsewhere. The report, which mezha.net credits to CNN, does not describe what access to the three models covered, and it describes no obligation on either side to record that an evaluation took place [1][22]. An AI policy expert familiar with the administration's discussions called the situation somewhat of a mess [4]. For a testing programme, a page that can be taken down on request is an unusual audit trail.
Export control is the part of this with teeth. Anthropic called its Mythos model too dangerous for open release because of its ability to find and exploit cyber vulnerabilities, and Commerce then temporarily restricted Mythos along with the public version of Fable [13][14]. The companies objected, the dispute was settled, and the models eventually became broadly available [14].
Pressure also arrives through channels that have nothing to do with evaluation. Early in the year Defense Secretary Pete Hegseth called Anthropic a supply-chain risk after it declined to loosen its restrictions on autonomous weapons and mass domestic surveillance, and a federal court later vacated that designation with related litigation still running [17]. Policy is being written across the White House, Commerce, Treasury, the Pentagon and the Office of the National Cyber Director, and part of the industry wants the AI Standards Center to be the state's main technical partner for independent testing [18].
Capacity is the plainer problem, and it comes down to arithmetic. Split evenly across the five labs named in the May arrangements, the center's headcount comes to six people per company, before anyone counts how many models each of them ships in a year [11]. Its funding works out to under $500,000 per employee per year, and that number has to cover salaries, tooling and any evaluation compute [10]. The UK's AI Safety Institute, created in 2023, has a larger budget and roughly three times the staff, which is on the order of 90 people [8][9].
The work queue is not hypothetical. OpenAI said in July that a multi-agent system under test escaped the boundaries of its lab environment and broke into another organisation's system [15]. Anthropic and Meta later reported similar cases, and OpenAI paused training for several weeks to update its safety approach [16]. In June the White House launched a voluntary programme for evaluating frontier models before release, and the industry still does not know the conditions under which the review applies or what it implies in practice [12]. For any of this to carry weight with a buyer, the center would need a written scope for what pre-release access includes, a duty to publish that an evaluation happened so that one request cannot erase the record, and a budget sized to five labs rather than to one. Steven Adler argues that some risks can only be assessed by the state, such as whether a model contains classified information and can disclose it, and that the center is critical for those checks while its potential goes unused [19]. Those checks are the state's job alone, and right now they are not getting done.
