“We want access to all the models, and we’re going to take advantage of that,” Tim Kosiba, the National Security Agency’s deputy director, told an Intelligence and National Security event in Bethesda, Maryland.
Tim Kosiba and the NSA’s push for broad model access
Kosiba’s remark is one of the clearest public declarations that the NSA is actively engaging commercial developers as it implements new White House-directed processes for evaluating advanced artificial-intelligence systems. He said the agency has used AI in various forms for decades but that “the transformative change that we see today is what these models can actually do,” prompting a more aggressive effort to obtain and deploy them.
Kosiba also made explicit limits to what he disclosed: he did not identify which companies the NSA is talking to, did not name the models the agency currently uses, and did not say whether any developer has provided the agency access to an unreleased system.
The June executive order, the voluntary testing window, and agency roles
The remarks come as the NSA helps implement a June executive order that directs national security agencies to develop classified tests for when AI systems have acquired sufficiently advanced hacking capabilities to warrant extra scrutiny. The order specifies a voluntary arrangement under which developers can give the federal government access to qualifying models for as many as 30 days before releasing them to other trusted partners, while explicitly prohibiting a mandatory government licensing or approval process for new models.
The NSA director — in consultation with the Office of the National Cyber Director, the Cybersecurity and Infrastructure Security Agency, and other government offices — is responsible for deciding which systems meet the threshold for a “covered frontier model.” A related national security directive from June instructs intelligence and defense agencies to form “deep, proactive partnerships” with industry and to make the most advanced models available to national security personnel without delay, while encouraging diversification of suppliers to avoid dependence on a single company.
This site is the portfolio.
OSINTSights runs on Cloudflare Workers, D1, R2, and Vectorize, with an AI pipeline on Hetzner ARM. Nubivance designed, built, and operates it. We do the same for clients.
See what we buildWhat “access” can mean, and what the record shows so far
- The term “access,” Kosiba acknowledged, can cover several arrangements: using a model through a company-controlled service, deploying a version inside a government environment, or receiving access for security testing of classified networks.
- Kosiba stopped short of saying the voluntary pre-release program is fully operational or that companies have begun providing models through it.
- The White House told major developers earlier this month that open-weight models — those whose underlying settings can be downloaded and modified — would not be included in the testing program.
- Separately, OpenAI’s latest GPT-5.6 models “became available to federal customers this month through its FedRAMP-authorized enterprise service after their public release in July,” the reporting notes, an example of expanded commercial access to government customers.
Anthropic, Mythos, and the Defense Department dispute
The push for broad government access is unfolding alongside a contested relationship between the Defense Department and Anthropic. The Pentagon designated Anthropic a supply-chain risk after the company declined to relax some restrictions on certain military uses of its models; that designation prompted a lawsuit in which a federal judge temporarily blocked parts of the government’s action. The NSA, according to reporting in The New York Times cited in the record, has secured carveouts with Anthropic and is “already using versions of Anthropic’s advanced Mythos model on an experimental basis” to test defenses and to evaluate the model’s ability to find weaknesses in foreign networks.
Those arrangements have been subject to policy friction: parts of the NSA temporarily lost access to Mythos 5 amid an export-control dispute between the company and the previous administration, and the administration later lifted the restrictions that had affected access.
What this means for technologists, policymakers, and the Defense Department
- Technologists and security teams: Expect government engagement in testing and evaluation to take multiple technical forms — cloud-based access, in‑government deployments, and targeted security testing — each carrying different operational constraints and audit requirements, as Kosiba noted without specifying which approach will be used in particular cases.
- Policymakers and regulatory officials: The executive order’s explicit ban on a mandatory licensing regime creates a voluntary, time-limited pathway (up to 30 days pre-release) that places judgment about “covered frontier models” in the hands of the NSA director in consultation with ONCD and CISA.
- The Defense Department and military acquisition officials: The Anthropic dispute illustrates that company-imposed usage restrictions and export-control decisions can suddenly alter government access — a dynamic that already removed and later restored some NSA access to Mythos 5.
Kosiba emphasized one final operational constraint: regardless of how quickly the NSA adopts AI, human reviewers will remain responsible for ensuring compliance with surveillance laws and internal rules. “What the models are putting out, how we’re leveraging technology today, will always require that human to double-, triple-check and make sure we are in complete compliance with the law,” he said.
The facts on the table show an agency seeking wide-ranging entry points into commercial models while operating under a voluntary White House framework. What remains unsettled in the public record is whether the voluntary pre-release program is yet accepting models, which developers — if any — have agreed to participate, and how the government will reconcile company usage constraints, export controls, and legal challenges as it seeks access to rapidly evolving systems.
