US government access to AI models secured from Microsoft, Google and xAI for security testing
Microsoft, Google and xAI grant US government access to AI models for security testing as CAISI ramps evaluations amid concerns over Anthropic’s Mythos
The United States has secured government access to new AI models from Microsoft, Google and xAI to conduct national-security testing, officials said on Tuesday. The agreement gives federal scientists the ability to evaluate and probe the systems before they are broadly deployed, aiming to identify risks to cybersecurity and military use. US government access to AI models will allow the Center for AI Standards and Innovation to run targeted tests and research workflows designed to expose unexpected behaviors. The move follows recent advances in generative AI that have heightened concern among policymakers and intelligence officials.
Agreements with Microsoft, Google and xAI
Microsoft, Google and xAI agreed to let US government teams test their latest frontier models under the new arrangement announced by CAISI. Company representatives will collaborate with government scientists to develop shared datasets and testing procedures that simulate misuse scenarios and probe system vulnerabilities. Microsoft said it will work with US researchers to examine “unexpected behaviors” and refine evaluation methods, while Google confirmed it would cooperate with testing protocols. xAI’s response was not immediately available at the time of the announcement.
CAISI expands role in federal model testing
The Center for AI Standards and Innovation, part of the Commerce Department, will be the central hub for evaluations under the agreement. CAISI has broadened its remit since its earlier incarnation and reports it has completed more than 40 model evaluations, including analyses of systems not yet publicly released. Director Chris Fall emphasized that independent, rigorous measurement science is essential to understanding frontier AI and its national-security implications. The center will accept developers’ models, often with certain safety guardrails reduced, so researchers can stress-test potential failure points.
Context of prior agreements and personnel moves
The new pact builds on earlier arrangements forged under the previous administration when the institute operated under a different name and mandate. CAISI previously reached agreements with major AI developers to create voluntary safety standards, tests and definitions intended to measure system behavior. Officials noted that some staff who led testing efforts in prior years subsequently joined industry, including a senior adviser who is now at Anthropic, according to publicly available profiles. That movement highlights the continuing interchange between government and private-sector AI expertise.
Concerns driven by Anthropic’s Mythos and broader threats
Recent attention has focused on Anthropic’s newly unveiled Mythos model, which US officials and corporate security teams say could enhance the capabilities of malicious actors. The administration and CAISI officials voiced concerns that powerful models might be repurposed to accelerate cyberattacks, automate misinformation campaigns, or assist in weaponization scenarios. By obtaining early access to frontier models, the government aims to surface these risks before models are widely distributed. CAISI will use tailored evaluations to estimate attack vectors and inform mitigation strategies.
Pentagon pact with seven tech firms and Anthropic’s absence
The announcement followed a separate Department of Defense agreement with seven large tech companies to deploy their AI systems on classified networks to augment warfighter decision-making. The Pentagon said those partnerships would help integrate commercial AI tools into complex operational environments under classified conditions. Notably, Anthropic was not included in the Department of Defense list after a public dispute and legal contest with the administration over the rules governing AI use in military contexts. The absence highlights rising tensions between some AI developers and federal defense agencies.
Market reaction and industry implications
Financial markets showed a mixed response to the news: Microsoft shares dipped modestly in midday trading while Alphabet’s parent company, Google, saw gains, and xAI remains privately held. Analysts said investor moves reflect a balance between potential regulatory scrutiny and the commercial value of close ties to government testing programs. For tech firms, offering controlled access to models can help shape standards and demonstrate a commitment to safety while still enabling rapid product development. Observers cautioned that these pacts may prompt calls for clearer statutory safeguards and oversight of model deployment.
The agreement to grant US government access to AI models from major developers marks a significant step in federal oversight of advanced systems, reflecting heightened concern about the national-security implications of frontier artificial intelligence. As CAISI scales its evaluations and the Pentagon formalizes partnerships with industry, regulators, companies and researchers will face pressure to translate testing findings into enforceable safeguards and operational protocols. The coming months are likely to reveal how effectively coordinated testing can limit misuse while allowing beneficial innovation to proceed.
