Public Policy & Governance

The Defense Department officially designated Anthropic a supply-chain risk after the company refused contract terms requiring the Pentagon to control how its AI models could be used, including for autonomous weapons and domestic surveillance. According to Bloomberg, this is the first time a US company has received this classification, which until now has only been applied to companies from adversary nations like China's Huawei. The designation followed Anthropic's rejection of a $200 million contract; the DoD immediately pivoted to OpenAI, which accepted similar terms. Multiple sources report that Anthropic amended its core safety principle amid the dispute, though the company maintains it will not support military surveillance or weapons applications.

The designation carries significant implications beyond the immediate contract. As TechCrunch reports, Microsoft, Google, and Amazon confirmed that Claude remains available to non-defense customers through their cloud platforms, but the supply-chain risk label could trigger reviews across other federal agencies. Heidy Khlaaf, chief AI scientist at AI Now Institute, told multiple outlets that existing guardrails for generative AI are 'deeply lacking' and 'easily compromised,' questioning whether any company can credibly promise controls sufficient for military applications. The dispute exposes a fundamental tension: federal procurement practices assume vendors will customize products to government specifications, but AI companies argue their safety frameworks cannot be overridden without compromising the entire model architecture.

The US Commerce Department has drafted regulations that would require government permits for Nvidia and AMD AI chip shipments to anywhere in the world, according to Bloomberg reporting. This represents a dramatic expansion from current targeted controls focused on China and a handful of other countries. The draft regulations would establish a universal licensing regime where any export of advanced AI semiconductors requires explicit American approval, effectively treating all countries as potential diversions risks. The timing coincides with heightened concerns about Gulf region data centers becoming conflict targets, as noted by Carnegie Endowment fellow Sam Winter-Levy in Bloomberg coverage of Iran conflict risks to regional computing infrastructure.

The regulatory draft follows months of semiconductor industry lobbying against broader controls, with companies arguing that universal licensing would cede markets to non-US competitors. However, the Iran conflict has shifted the calculus for national security officials who now view data center concentrations in geopolitically unstable regions as strategic vulnerabilities. The draft includes no exemptions for treaty allies, which will likely trigger diplomatic pushback from EU and Asian partners who have their own AI sovereignty initiatives. Implementation details remain unclear, particularly regarding what constitutes an 'advanced' AI chip subject to licensing and how quickly Commerce can process what could become tens of thousands of permit applications.

Draft federal procurement regulations would require civilian government AI contracts to include terms making models available for 'any lawful use', according to Financial Times reporting. The guidelines directly address the Pentagon-Anthropic impasse by establishing baseline terms that prevent AI companies from imposing use restrictions beyond what statute requires. The draft language was developed by the Office of Management and Budget in consultation with the General Services Administration and would apply to civilian agencies including the Department of Homeland Security, Health and Human Services, and others pursuing AI adoption. Unlike defense contracts, which can invoke national security authorities to mandate specific access terms, civilian procurement has historically relied on commercial-off-the-shelf software agreements where vendors retain significant control over acceptable uses.

The draft guidelines would effectively override vendor policies that restrict government use of AI models for surveillance, decision-making about benefits eligibility, or other applications that companies consider high-risk. This reverses the current dynamic where companies like Anthropic have been able to decline government business based on their internal acceptable use policies. Legal experts note that 'any lawful use' is deliberately broad language that would permit agencies to deploy AI for purposes that may be legal but constitutionally questionable absent clear statutory prohibition. The guidelines are still in draft form and have not been released for public comment, suggesting OMB is coordinating with affected agencies before formal rulemaking.

The Cybersecurity and Infrastructure Security Agency added three iOS vulnerabilities to its catalog of known exploited flaws, according to Ars Technica, which describes 'a long, strange trip of a large assembly of advanced iOS exploits' detected under mysterious circumstances. CISA's Known Exploited Vulnerabilities catalog triggers mandatory patching requirements for federal agencies under Binding Operational Directive 22-01. The designation indicates CISA has evidence the vulnerabilities are being actively exploited in the wild, though the agency has not disclosed attribution or targeting details. The timing suggests federal concern about mobile device security extends beyond theoretical risks to active compromise campaigns affecting government personnel or contractors.

The catalog addition comes as federal agencies accelerate mobile-first strategies for remote work and field operations, creating expanded attack surfaces. CISA's decision to publicly flag iOS exploits is notable given the agency typically focuses on server and network infrastructure vulnerabilities. The lack of public attribution, combined with Ars Technica's description of 'advanced' exploits deployed under 'mysterious circumstances', suggests sophisticated state-sponsored activity rather than commodity malware. Federal agencies now have limited time to ensure all iOS devices are updated, though the practical challenges of enforcing mobile patching across distributed workforces remain significant.