White House Mulls Pre-Release AI Model Reviews

The administration is discussing an executive order to establish an AI working group and explore a formal government review process for new models before public release, reporting by The New York Times says. The New York Times reports that senior officials briefed executives from Anthropic, Google, and OpenAI about some of the deliberations. Multiple outlets link the discussions to Anthropic's new model Mythos, which reporting describes as having advanced coding and vulnerability-finding capabilities that have raised cybersecurity concerns. Reuters reports a White House official said, "Any policy announcement will come directly from the president. Discussion about potential executive orders is speculation." Reuters and The New York Times note that the reported deliberations contrast with President Trump's earlier revocation of a 2023 executive order from the Biden administration that had required certain AI safety test reporting.

Industry-pattern observations: governments weighing pre-release reviews are most often responding to models that increase offensive cyber capabilities or rapid dual-use risk, per the coverage linking Mythos to potential exploitation of software vulnerabilities (Reuters, The New York Times, Economic Times). The UK example is cited in The New York Times as a reference model for assigning government bodies to certify that models meet safety standards. For practitioners, pre-release review regimes typically raise friction for continuous-delivery workflows and increase the importance of robust internal red-teaming, documented safety tests, and reproducible evaluation artifacts.

public reporting frames the deliberations as a policy reversal relative to the administration's earlier deregulatory stance, creating legal and compliance uncertainty for model developers. Companies that produce advanced models may face new administrative steps before public deployment if an executive order or comparable rules are adopted. The citing of Mythos in multiple reports illustrates how a single high-capability model can trigger broader policy attention when journalists and officials highlight dual-use risks such as automated vulnerability discovery.

• •Whether an executive order is formally issued and the precise statutory or administrative authority it cites, which will determine legal enforceability (reporting so far is preliminary).
• •Which government entities would participate in a working group and whether the process would be advisory, mandatory, or tied to funding and export controls (The New York Times reports the working group concept without implementation details).
• •How major developers document and share safety evaluations, since prior U.S. policy addressed sharing safety test results under the 2023 Biden executive order (Reuters reporting).

For practitioners: engineering and security teams should monitor official guidance and be prepared to improve auditability of model training, testbeds for red-team results, and provenance records. Industry observers note that comparable oversight regimes increase demand for standardized evaluation artifacts and third-party audit tooling.

All reporting so far describes deliberations rather than finalized policy. A White House official characterized talk of an order as "speculation," Reuters reports. The New York Times and Reuters cite unnamed U.S. officials and people briefed on the discussions, and neither outlet reports a signed directive at the time of publication.