IC Technology & Surveillance — 2026-05-20

Executive Order Expected to Give NSA Role in Voluntary AI Model Safety Testing

BLUFNSA will likely secure the lead evaluator role before September 1, 2026, but the voluntary architecture leaves the arrangement an intelligence access channel rather than a binding safety regime.

White House officials are planning a voluntary information-sharing framework between the government and AI developers for pre-deployment safety testing, with the NSA expected to conduct classified model evaluations before public release, according to multiple unnamed sources cited by Nextgov 1. Bloomberg reported on May 8 that the forthcoming order omits mandatory testing requirements 2. Nextgov sources identified Anthropic's Mythos Preview and OpenAI's GPT-5.5, both capable of network vulnerability discovery, as principal catalysts for the deliberations 1. Nextgov also cited a Washington Post report finding spy agencies and the Commerce Department at odds over which entity should lead model evaluation 1.

Analysis
An executive order naming NSA as primary evaluator of advanced AI models will likely be signed before September 1, 2026, resolving the interagency contest with Commerce in the intelligence community's favor, corroborated without contradiction across three outlets. The voluntary architecture is the order's defining constraint: labs retain the right to decline, making this an access channel rather than a regulatory checkpoint. That voluntary design may equally reflect industry lobbying success as genuine security concern, providing symbolic oversight while leaving capable labs free to deploy without evaluation. If signed, AI labs face an immediate binary: submit pre-deployment models to NSA classified review or decline.

4 sources