The US Department of Defense (referred to in some 2026 releases as the Department of War) is the largest single government consumer of AI services. Its AI adoption spans intelligence analysis, logistics, battlefield decision-making, autonomous systems, and administrative modernization. Through 2025 and 2026 the department expanded both unclassified and classified AI deployment while becoming the focal point of a governance dispute over which uses of commercial AI models — particularly mass surveillance and autonomous weapons — vendors would permit.
Organization and AI offices
- CDAO (Chief Digital and AI Office) — DOD's principal AI office, headed by Radha Plumb.
- Defense Innovation Unit (DIU) — commercial-technology pipeline.
- AI Rapid Capabilities Cell (AI RCC) — a CDAO and DIU collaboration for frontier-AI pilots.
- Task Force Lima — DOD's generative AI task force. See CSET + DOJ + CSIS — US-China AI National-Security Axis (composite source summary).
GenAI.mil and unclassified deployment
The Department of Defense's GenAI.mil platform is the central conduit through which DOD personnel access frontier models. On April 27, 2026, the Pentagon added Google's latest AI model to GenAI.mil, with platform usage reported as rising (Source: nextgov.com).
On April 28, 2026, the Pentagon disclosed that users had built more than 100,000 AI agents on GenAI.mil and that Google's latest model had been added to the available roster (Source: defenseone.com). Reporting the same week put the figure higher, stating that staff had used Gemini Agent Designer to build more than 103,000 semi-autonomous AI agents on unclassified networks in under five weeks (Source: breakingdefense.com). This unclassified deployment predates the May 1 classified-networks agreements, which extend rather than initiate federal Gemini agent use. See AI Coding Agents, Google DeepMind.
ChatGPT Mil became the platform's second approved generative-AI tool for unclassified work on August 31, 2026. A Department release described it as bringing "a familiar commercial experience into the Department's secure environment, tailored to warfighter needs" and supporting "document-heavy unclassified work across the Department, including planning, policy, logistics, and administration," built to scale to more than 3 million Department personnel. Work on a version of ChatGPT usable with controlled unclassified information began in February 2026, and an OpenAI official said in June that deployment was expected in "early July." A defense official speaking anonymously ahead of the announcement said the model is "engineered in a way to be separate from the commercial model"; the Defense Information Systems Agency builds the environment and the National Security Agency probes it for ways an attacker might invade it or the model might escape. Joe Larson, who leads OpenAI's government sales, said the integration is "grounded in our national security principles, pairing access to powerful tools with clear safeguards, meaningful human judgment, and accountability for high-stakes decisions." Officials from both the Department and OpenAI said in background discussions that they do not want GenAI.mil models used for combat (Source: defenseone.com).
GenAI.mil launched on December 9 with Google's Gemini as its only tool. Pentagon officials signed an agreement around the same time to add xAI, which has not been formally deployed (Source: defenseone.com).
A separate, non-combat application drew attention earlier: WIRED reported that the US Army was using "CamoGPT," an LLM-based content-filtering tool, to purge diversity, equity, and inclusion material from training materials.
Classified-networks procurement
The Department of War announced on May 1, 2026 that it had finalized agreements with eight AI model and infrastructure companies — SpaceX, OpenAI, Google, Nvidia, Reflection, Microsoft, AWS, and Oracle — to deploy frontier AI on classified networks "for lawful operational use." The agreements excluded Anthropic over the company's refusal to permit mass domestic surveillance and fully autonomous weapons targeting (Dario Amodei's "two non-negotiables"). Reflection was identified as the first open-weight model developer signed for classified work. Former White House AI czar David Sacks criticized Anthropic publicly, saying the company "tried to set themselves up in a way of being superior to the chain of command." The exclusion ran alongside parallel reporting that the White House was drafting an executive workaround to allow federal agencies to onboard Mythos despite its supply-chain-risk designation, and a separate national-security AI memo preparing to require federal agencies to use multiple AI vendors and to "prohibit AI providers from interfering with the military's chain of command" (Sources: bloomberg.com, reuters.com, politico.com, insideaipolicy.com, axios.com, bloomberg.com). See Anthropic Claude Gov + Pentagon Dispute (2025–2026), Anthropic, OpenAI.
The DoD-released agreement text was clarified on May 2, 2026 to specify that the deals covered IL6 and IL7 networks — classified up to TS/SCI per the DoD Cloud Computing Security Requirements Guide (Source: war.gov).
The department is seeking congressional authority to redirect $600 million toward standing up a high-performance AI compute centre at the Top Secret/Sensitive Compartmented Information level, according to a budgetary reprogramming package obtained by Inside Defense/AI and reported on September 8, 2026. The funds would procure computing infrastructure, hardware and integration services across the Joint Worldwide Intelligence Communications System (Source: insideaipolicy.com). The reporting rests on a scrape that returned only the article's lede, so the request's status in Congress and its relationship to the classified-network agreements above are not established here.
May 4 reporting reframed the cohort as seven companies — SpaceX's xAI, OpenAI, Google, Nvidia, Reflection, Microsoft, and AWS — with Anthropic still excluded and Oracle no longer named, superseding the earlier eight-vendor listing that had included Oracle. Reflection remained the only open-weight developer in the cohort (Sources: theinformation.com; insideaipolicy.com; apnews.com; washingtonpost.com).
Sen. Elizabeth Warren (D-MA) sent letters — made public July 9, 2026 — to Defense Secretary Pete Hegseth and executives at Google, SpaceX, NVIDIA, Reflection AI, Microsoft, Amazon Web Services, and Oracle, requesting the full text of the May commercial-AI agreements and details of the safeguards governing military use (Source: insideaipolicy.com).
On the infrastructure side, SpaceX's talks to provide billions of dollars' worth of data-center capacity for the Pentagon's AI buildout became public on July 17, 2026 (Source: wsj.com). See xAI.
Google formally signed an agreement with DoD on April 27, 2026 allowing Pentagon use of Google's AI models on classified work for "any lawful government purpose" — the same contractual language Anthropic refused and OpenAI accepted. Google proposed contract language excluding lethal autonomous weapons and mass surveillance; civil-liberties lawyers described the language as insufficient. The Google deal was the third frontier-lab-to-Pentagon contract under the "any lawful use" framework, after OpenAI (March 2026) and Anthropic's rejection of the same terms (Source: theinformation.com).
Vendor relationships
| Vendor | Scope | Source |
|---|---|---|
| Anthropic (Claude Gov) | Classified environments; intelligence analysis; cyber | Anthropic Claude Gov + Pentagon Dispute (2025–2026) |
| Palantir | AIP, Foundry, Gotham; $10B US Army contract (July 31, 2025) | Washington Post |
| Anduril | Defense AI; autonomous systems |
Employee opposition
More than 600 Google employees, including many from DeepMind, signed a letter to CEO Sundar Pichai on April 27, 2026 demanding that he bar the Department of Defense from using Google's AI for classified work, arguing that "the only way to guarantee that Google does not become associated with such harms is to reject any classified workloads." The letter followed the GenAI.mil expansion announcement and was the most public employee challenge to a US frontier lab's military-cloud relationship since the Project Maven walkout of 2018 (Source: washingtonpost.com).
The Anthropic–Pentagon dispute
The dispute over the terms on which the Pentagon could use Anthropic's models began in early 2026:
- A Hegseth memorandum dated January 9, 2026 and posted January 12, Artificial Intelligence Strategy for the Department of War, directed that standard "any lawful use" language be incorporated into any departmental contract procuring AI services within 180 days, and argued that "the risks of not moving fast enough outweigh the risks of imperfect alignment." The memorandum is the originating text of the dispute; The Verge surfaced it on May 26, 2026 and dated it January 12, the posting date, while the signed text carries January 9 (Artificial Intelligence Strategy for the Department of War (Hegseth memorandum, 2026)).
- The Pentagon sought "all lawful use" of Claude, including for autonomous weapons and mass surveillance.
- Dario Amodei refused, citing two non-negotiables on the commercial-product side; he has separately written that "fully autonomous weapons … may prove critical for our national defense" on the research-and-development side.
- Secretary Hegseth announced on February 27, 2026 that Anthropic would be designated a supply chain risk; the formal written determinations followed on March 3, 2026, one under 10 U.S.C. § 3252 reviewed in the Northern District of California and one under the Federal Acquisition Supply Chain Security Act, 41 U.S.C. § 4713, reviewed in the D.C. Circuit (Anthropic PBC v. U.S. Department of War — Order on Cross Motions for Summary Judgment (N.D. Cal., August 27, 2026); Anthropic PBC v. U.S. Department of War — D.C. Circuit Opinion, Nos. 26-1049 and 26-1162 (September 25, 2026); Federal Acquisition Supply Chain Security Act of 2018 (FASCSA)).
- Judge Lin issued a preliminary injunction blocking the ban on March 26, 2026.
- At D.C. Circuit oral arguments on May 19, 2026, Judge Henderson called the supply-chain-risk designation "spectacular overreach."
- Summary judgment for Anthropic on August 27, 2026. The court held the measures unlawful First Amendment retaliation and a due-process violation, vacated the designation and the contractor-boycott order under the Administrative Procedure Act, vacated nine agencies' implementing orders, and entered a permanent injunction (Anthropic PBC v. U.S. Department of War — Order on Cross Motions for Summary Judgment (N.D. Cal., August 27, 2026); Anthropic PBC v. U.S. Department of War — Order of Final Relief (N.D. Cal., August 27, 2026)).
- Reaffirmed by the Department on September 3, 2026. Under Secretary of Defense for Research and Engineering Emil Michael posted on X that "Anthropic is still a designated Supply Chain Risk at @DeptofWar and for the Defense Industrial Base," a day after Commerce Secretary Howard Lutnick said the company was "back on the right side" and had resolved its issues with the administration (Source: axios.com; bloomberg.com). The statement postdates the August 27 summary judgment vacating the designation, and Michael is the same Under Secretary whose office the court held could not supply the required risk assessment. Anthropic continues to contest a separate Pentagon designation under a different statute in the D.C. Circuit (Source: axios.com).
- D.C. Circuit upholds the § 4713 exclusion on September 25, 2026. A divided panel (Katsas, J., joined by Rao, J.; Henderson, J., dissenting) denied Anthropic's petitions for review of the Department's parallel exclusion under the Federal Acquisition Supply Chain Security Act, 41 U.S.C. § 4713, holding that the Department "had ample support for its conclusion that the continued integration of Claude into the Department's information systems, by the Department or its contractors, presented a statutorily covered national-security risk" and that § 4713, unlike 10 U.S.C. § 3252, requires no bad motive (Anthropic PBC v. U.S. Department of War — D.C. Circuit Opinion, Nos. 26-1049 and 26-1162 (September 25, 2026)). The majority found the Department reasonably feared that Anthropic "might manipulate Claude's design to prevent it from performing national-security functions that the Department deems contractually authorized and necessary," and that "the Department cannot utilize AI systems that remain trapped in amber"; it accepted that the Chief Information Officer's March 6, 2026 removal order set 180 days only as an outer bound reflecting the difficulty of removing the technology "in the midst of active operations" (Anthropic PBC v. U.S. Department of War — D.C. Circuit Opinion, Nos. 26-1049 and 26-1162 (September 25, 2026)). CNBC reported that the panel delayed the decision's effect to allow Anthropic to seek rehearing (Source: cnbc.com). See Anthropic v. United States (Pentagon ban challenge).
The judgment records the exclusion authority's limits. The court read 10 U.S.C. § 3252 as "directed at covert acts or hacks, not overt, public positions regarding contract terms," and held that the Department had failed the statute's own safeguards: the written determination and congressional notification about less intrusive measures required by § 3252(b)(2)(B) and (b)(3)(B), and the requirement in 48 C.F.R. § 239.7304(a) that the joint recommendation rest on a risk assessment by the Under Secretary of Defense for Intelligence rather than, as here, the Under Secretary for Research and Engineering then negotiating the contract. It also recorded that the Department conceded no statutory basis existed for the order barring military contractors from any commercial activity with Anthropic. The order leaves the Department free to select another vendor and does not require it to use Anthropic's products (Anthropic PBC v. U.S. Department of War — Order on Cross Motions for Summary Judgment (N.D. Cal., August 27, 2026); Anthropic PBC v. U.S. Department of War — Order of Final Relief (N.D. Cal., August 27, 2026)).
The opinion also documents the pre-dispute procurement record: Claude Gov in use through partner platforms since March 2025, Top Secret facility clearance after an 18-month vetting process, FedRAMP High and DoD Impact Level 4 and 5 authorization in June 2025, and a two-year CDAO agreement worth up to $200 million in July 2025, with no supply-chain-risk concern identified before March 2, 2026 (Anthropic PBC v. U.S. Department of War — Order on Cross Motions for Summary Judgment (N.D. Cal., August 27, 2026)).
At a Senate hearing on April 29, 2026, Defense Secretary Pete Hegseth said Anthropic CEO Dario Amodei was "an ideological lunatic who shouldn't have a sole decision-making over what we do" (Sources: platformer.news, insideaipolicy.com).
Around April 29, 2026 the White House was reported to be workshopping a plan to restore Anthropic's role in federal AI contracts, reversing the earlier Trump-administration characterization of Anthropic as a "supply chain risk." The plan would re-open Pentagon and intelligence-agency access to Claude Mythos's cyber-vulnerability detection (Source: axios.com).
Emil Michael, the department's chief technology officer, said in an account published September 11, 2026 that approximately 90 percent of the classified AI workloads that had run on Anthropic technology had been transitioned to other model providers, and that all of the firm's technology in the Maven Smart Systems program had been transitioned out months earlier. The transitions proceeded despite a federal judge's ruling in late August 2026 that the department's designation of Anthropic as a supply chain risk earlier in the year was "illegal and baseless." Only the lede of the trade-press account was retrievable, so the identity of the replacement providers and the status of the remaining 10 percent are not established (Source: insideaipolicy.com). See Anthropic v. United States (Pentagon ban challenge).
Full analysis: Anthropic Claude Gov + Pentagon Dispute (2025–2026), Anthropic v. United States (Pentagon ban challenge), Pete Hegseth, Clawed.
Intelligence-community use of Mythos
The NSA had been testing Anthropic's Mythos to find vulnerabilities in Microsoft products and other widely used software — the first publicly reported intelligence-community use of Mythos's cyber-vulnerability-discovery capability, reported April 30, 2026. The NSA testing predated and ran parallel to the public Pentagon-Anthropic conflict, indicating the intelligence community retained access through a separate channel (Source: bloomberg.com). See National Security Agency (NSA).
By June 5, 2026, reporting stated that the NSA was using Mythos to conduct offensive cyber operations, possibly with Anthropic staff embedded at the agency — a shift from the earlier vulnerability-testing (defensive) framing to active offensive use, and a contrast with the Pentagon's contested supply-chain-risk exclusion of Anthropic (Source: ft.com). See National Security Agency (NSA), Anthropic.
Autonomous systems and special operations
On April 28, 2026, SOCOM commander Gen. Bryan Fenton said that Special Operations Command was adding AI and autonomy "at every level" of its operations (Source: defenseone.com). Special-operations integration of AI proceeded even where the frontier-lab supply chain was contested, a counterpart to the Anthropic Pentagon dispute over autonomous-weapons restrictions.
The same week, Hegseth disclosed that the Pentagon would soon establish a sub-unified command for autonomous drone warfare, formalizing the SOCOM "AI/autonomy at every level" framing into command-structure changes (Source: insideaipolicy.com).
On April 28, 2026 — the same period in which Google signed the classified Pentagon AI deal — Google withdrew from a $100M Pentagon voice-controlled drone-swarm contest after advancing through earlier rounds, citing an internal ethics review (Source: bloomberg.com). See Autonomous Weapons.
The Air Force published its Department-wide data and AI strategies on April 29, 2026, focused on accelerating decision-making capabilities — the first Department-wide AI strategy publications from any service since the America's AI Action Plan (Source: insideaipolicy.com).
The Defense Innovation Unit issued a solicitation reported May 4, 2026 for paired modular data centers — one semi-fixed, one mobile — to support AI-enabled operations in contested battlefield environments, extending AI infrastructure to the tactical edge (Source: insideaipolicy.com). See Defense Innovation Unit (DIU).
At the annual AWS summit on June 30, 2026, Pentagon Chief Digital and AI Officer Cameron Stanley said a new "Agent Network" and "war data platform," built with AWS, Palantir, and Anduril, are enabling near-instant kill-chain targeting decisions, with officials seeking to convert the pilot into a program of record (Source: insideaipolicy.com). See Autonomous Weapons.
A vendor-side view of the adoption curve came from Shield AI president and co-founder Brandon Tseng, who said in an interview reported June 24, 2026 that while the department has rapidly adopted generative AI for workload and decision-support tasks, it remains in an "early adopter" phase for battlefield autonomy and physical AI (Source: insideaipolicy.com). The framing matches the pattern visible across DOD AI procurement: analytic and back-office generative-AI tools have diffused faster than autonomous physical systems, which face higher reliability, doctrine, and autonomous-weapons-policy hurdles.
Ground autonomy nonetheless has an operational track record. Reporting on July 21, 2026 detailed a $114 million US Army program backing semiautonomous breaching vehicles — already deployed by Ukraine's military against Russian fortifications — with program lead Sgt. Maj. Corey Wilkens saying future versions could be "completely agentic" (Source: puck.news). See Autonomous Weapons.
AI in targeting and intelligence analysis
In a memorandum dated January 9, 2026, "Artificial Intelligence Strategy for the Department of War," Defense Secretary Pete Hegseth directed the department to become an "AI-first" warfighting force and to make AI available across it, "putting America's world-leading AI models directly in the hands of our three million civilian and military personnel, at all classification levels" — the sixth of seven Pace-Setting Projects, GenAI.mil (Artificial Intelligence Strategy for the Department of War (Hegseth memorandum, 2026); Source: cnn.com). The memorandum also directs a vendor cadence deploying the latest models within 30 days of public release as "a primary procurement criterion," federated data catalogs at all classification levels delivered within 30 days, a monthly Barrier Removal Board empowered to waive non-statutory requirements, an AI Integration Lead named by each service chief and combatant commander within 30 days, and review by the Director of Cost Assessment and Program Evaluation "for resourcing adjustment" of exercises that do not meaningfully incorporate AI and autonomous capabilities. The strategy is described at Artificial Intelligence Strategy for the Department of War (2026). Two accounts published on September 18, 2026 describe errors in the use of AI for targeting and intelligence during the 2026 war with Iran.
The Minab school strike. Two Tomahawk missiles struck the Shajarah Tayyebeh Elementary School in Minab, Iran on February 28, 2026, killing more than 150 people including at least 123 children. Officials involved in the Pentagon's internal investigation, which Central Command commander Adm. Brad Cooper ordered in March and whose report has not been released, told Bloomberg that a cascade of preventable failures contributed: the main military intelligence database still catalogued the site as an Islamic Revolutionary Guard Corps facility, a classification the Defense Intelligence Agency had not updated although satellite imagery showed years of changes; the administration's demand for an overwhelming aerial assault, in which more than 1,000 Iranian targets were hit in the first 24 hours, compressed the time for finalizing targets; and some Central Command personnel relied too much on the AI in Palantir's Maven Smart System, expecting it to flag stale information or inconsistencies, for reasons that are not clear. The site was fed into Maven with its uncorrected classification and emerged as a recommended day-one target; target-list work that had taken hours in past conflicts was condensed into several minutes through Maven. No member of a civilian harm mitigation team reviewed the site, which the officials attributed both to staffing cuts, in which Hegseth reduced headcount across a number of civilian-harm teams by roughly 90% to fewer than 20 staff and Central Command's team from 10 to one, and to a decision not to involve the group in planning. Investigators concluded there had been high confidence throughout the kill chain that the site was an IRGC facility. A Palantir spokesperson said the company "is not responsible for the underlying data nor identifying intelligence deficiencies" and that there is no evidence its software was at fault; according to a person familiar with the matter, Palantir has since built capabilities into Maven to re-review underlying intelligence and flag inconsistencies. A Pentagon spokesperson said the incident remains under investigation (Source: bloomberg.com). The UN Independent International Fact-Finding Mission on Iran said on September 17, 2026 that there were reasonable grounds to conclude the strike amounted to a war crime (Source: news.un.org, as reported in bloomberg.com). On March 12, 2026 House Democrats led by Reps. Sara Jacobs, Yassamin Ansari and Jason Crow — more than 120 signatories, per Bloomberg — had put ten written questions to Hegseth about civilian harm in the Iran campaign, which the letter names Operation Epic Fury. The third asked what role AI played "in selecting targets, assessing intelligence, and making legal determinations," whether it is subject to human review and at what point, and whether AI "including the use of Maven Smart System" was used to identify the school as a target and whether a human verified it. The letter cited reports of over 1,000 total civilian deaths and asked whether the Department would investigate the school strikes as a possible war crime, and requested a response by March 20, 2026; the Pentagon has not answered such questions publicly, citing the investigation (House Democrats' letter to Secretary Hegseth on civilian harm in Operation Epic Fury (2026); Source: bloomberg.com).
U.S. Central Command's changes to its lethal targeting process after the strike became public on September 22, 2026. According to U.S. officials and people familiar with the matter, they refine how planned targets are refreshed and vetted, add open-source feeds to track civilian movement, and make dozens of upgrades to the Maven Smart System, including new ways to distinguish a building's function. Cameron Stanley, head of the Pentagon's digital and artificial intelligence office, said the upgrades include feeds tracking civilian maritime traffic; Stanley has said Maven enabled the 1,000 targets the United States struck on February 28 (Source: bloomberg.com). Use of the system grew during the war: James Mazol, a deputy under secretary of defense, said Maven's users more than doubled, from about 50,000 in January 2026 to more than 100,000, after U.S. attacks on Iran began under Operation Epic Fury, in remarks that became public on September 23; only the lede of that coverage was retrieved (Source: insideaipolicy.com).
A false intelligence report. In the spring of 2026 an intelligence report circulated across the US military stating that a Chinese ship in the Middle East was carrying components of a nuclear weapons program; the military planned to intercept the vessel, armed personnel prepared to board it and military aircraft were in the air, according to four sources. Shortly before the operation, officials found that a special operations command analyst had queried a chatbot about intelligence on the ship's manifest originating with US Special Operations Command Pacific, that the chatbot had fused open-source intelligence with classified signals intelligence and misidentified the cargo, and that the analyst had used AI again to package the finding as a standard intelligence report. One source called the report "entirely false" and said it "almost started a war." It was not clear whether the chatbot was a commercial product or a government one, and SOCOM Pacific and the Pentagon did not respond to requests for comment. Officials told CNN that AI adoption across the military is decentralized, with no single standard for verifying AI-generated information (Source: cnn.com). See Autonomous Weapons, Sycophancy and Hallucination.
The two episodes prompted a Senate request for Inspector General review. In a letter dated September 19, 2026 to Hegseth and Director of National Intelligence Jay Clayton, copied to Defense Department Inspector General Platte B. Moring III, Sens. Mark Warner (vice chairman of the Select Committee on Intelligence), Chris Coons (ranking member of the Appropriations Committee's Subcommittee on Defense) and Jack Reed (ranking member of the Armed Services Committee) cited the Minab strike and the aborted SOCPAC interdiction as evidence that the agencies had "prioritized acceleration of AI capability adoption and 'experimentation' over effective governance." They argued that these potential mistakes, "against the backdrop of the Department of Defense's harassment of a leading U.S. AI vendor over the latter's reluctance to allow its products' use in fully-autonomous kinetic operations," warranted immediate investigation by the relevant Inspectors General with greater public transparency in the findings. The letter contrasts the administration's revocation of National Security Memorandum 25 and a department AI strategy that "treats AI governance policies as 'blockers'" with AI governance measures Congress enacted in 2025 (Public Law 119-60) and OMB's April 2025 memorandum M-25-21, and asks that Inspectors General receive unrestricted access to both instances and any unreported ones (Sens. Warner, Coons and Reed — letter to Secretary Hegseth and DNI Clayton on AI targeting and intelligence errors (2026); Source: insideaipolicy.com). The senators' offices released the letter on September 21 (Source: warner.senate.gov). The letter does not name the vendor; Inside AI Policy, reporting it on September 24, identified the reference as the department's clash with Anthropic and described the senators as setting out potential oversight priorities should Democrats win a Senate majority in November (Source: insideaipolicy.com). An earlier House letter on Minab is summarized at House Democrats' letter to Secretary Hegseth on civilian harm in Operation Epic Fury (2026).
Other AI applications and procurement
U.S. Central Command announced on July 28, 2026 an agreement with the United Arab Emirates to establish Task Force Talon Synapse, described as the first bilateral task force for accelerating military AI applications. The Abu Dhabi-based team is to comprise roughly 20 American and Emirati personnel with AI, data and cybersecurity expertise (Source: centcom.mil). It is not yet operational: CENTCOM spokesperson Capt. Tim Hawkins said the first step is finalizing a memorandum of understanding, which CENTCOM is "working to sign … in the weeks ahead." A former senior defense official, speaking anonymously, called the announcement "unusually broad" (Source: defensescoop.com).
Strider, an intelligence firm using agentic AI alongside public records, expanded contracts with the US Air Force and NATO to identify foreign state actors in commercial supply chains, per Bloomberg reporting dated April 25, 2026 (Source: bloomberg.com).
Defense Department Undersecretary Emil Michael was, per a Washington Post profile dated April 27, 2026, deploying $200B in loans authorized by the recent reconciliation bill to defense-tech startups in biotech, AI, and minerals. Brookings' Darrell West warned that "the big problem is the lack of transparency in how they choose to work with private firms" (Source: washingtonpost.com).
A RAND report dated June 24, 2025, "Modernizing Department of Defense Civilian Human Resources: Harnessing AI for Transformative Change," examined applying AI to administration of DOD's roughly 700,000-person civilian workforce.
The Office of Personnel Management and the Defense Department launched "War Force," a Pentagon-targeted tech-engineer recruiting initiative under OPM's Tech Force program, on June 30, 2026 (Source: insideaipolicy.com).
The Navy adopted a strategy to "weaponize data and artificial intelligence" for faster decision-making and maritime operations, per a July 20, 2026 report (Source: insideaipolicy.com).
The department's commitment of more than $150 million across fiscal 2026–27 for a high-performance computing environment aligned with the Energy Department's Genesis Mission became public on July 23, 2026 (Source: insideaipolicy.com). See Science: A New Golden Age (OSTP, 2026).
Usage constraints have also surfaced: the U.S. Army told service members in an email, reported July 22, 2026, that they were "rapidly depleting their AI tokens" and needed to limit use (Source: wired.com).
Two AI vendor awards became public in August 2026. The AI company Raft said in an August 19, 2026 press release that the Army had awarded it a five-year, $99 million firm-fixed-price indefinite-delivery, indefinite-quantity contract establishing an "enterprise pathway" for acquiring and managing the company's software platforms, data integrations and "secure compute infrastructure," which Raft said gives the Army and the Defense Department a "streamlined path" to scale its capabilities without a new contract for each requirement. Only the lede of the trade-press account was retrievable, so the award date and the vehicle's ordering terms are not established here (Source: insideaipolicy.com). A Defense Department award of $100 million to the vendor NODA AI, to scale its AI-enabled mission command and coordination technology across the Joint Force, became public on August 18, 2026; no article body could be retrieved, and neither the award date nor the contract terms are verified (Source: insideaipolicy.com).
The Army launched Project Griffin on August 20, 2026 as a commercial solutions opening from the capability program executive for intelligence and spectrum warfare, seeking agentic artificial intelligence to process large volumes of sensor data and neutralize threats faster than humans can. Brandon Pugh described the initiative as part of a three-pronged effort within the Army Rapid Defensive Cyber Systems program (ARDS); the retrievable text does not state his title, and only the lede of the trade-press account was accessible, so the solicitation's terms, timeline and value are not established here (Source: insideaipolicy.com). See Autonomous cyber-agents, Agentic AI.
Compute capacity
Cameron Stanley, the department's chief digital and AI officer, told the National Defense Industrial Association's emerging technology conference in Washington on September 9, 2026 that the department needs significantly more computing capacity to accelerate military AI adoption, citing "a ridiculous amount of consumption across private industry and everywhere else." The department is seeking congressional authority to reprogram $600 million to establish a new classified AI compute center. Only the lede of the trade-press account was retrievable, so the proposed center's siting, timeline and relationship to existing classified enclaves are not recorded (Source: insideaipolicy.com). See Chief Digital and AI Office (CDAO), AI Infrastructure Capex.
In a separate strand of federal AI-infrastructure financing, the Pentagon was reported on September 10, 2026 to be in talks to lend roughly $5 billion to the AI cloud-computing startup Fluidstack through the Office of Strategic Capital, a lending office launched under the Biden administration for areas deemed critical to US national security. A loan of that size would be by far the office's largest to date. Fluidstack is being advised on the loan by the bank started by Palmer Luckey. The account rests on people familiar with the matter and on the retrievable lede of a paywalled report (Source: wsj.com).
Legislation and oversight
The House Armed Services Committee released the Chairman's mark of the FY27 NDAA on May 26, 2026, with markup scheduled for June 4, 2026. The bill contains AI provisions including updates to autonomous-weapons rules, accelerated deployment of AI models inside DoD, and incident-and-vulnerability reporting requirements (Source: insideaipolicy.com). See FY27 National Defense Authorization Act (Chairman's Mark).
On June 5, 2026, Sen. Elissa Slotkin introduced legislation to bar the Defense Department from using AI to surveil Americans or to launch nuclear weapons, targeting the 2027 NDAA as the vehicle. The bill would statutorily encode two of the red lines Amodei articulated as non-negotiables — no mass domestic surveillance and no fully autonomous launch authority — moving them from a vendor's contractual stance into proposed law. It pairs with Sen. Gillibrand's Secure and Accountable Military AI Act on the guardrails side of the military-AI debate (Source: notus.org). See FY27 National Defense Authorization Act (Chairman's Mark).
The US Cybersecurity and Infrastructure Security Agency (CISA), NSA, and counterparts in Australia, Canada, New Zealand, and the UK released joint guidance on responsible adoption of agentic AI on May 1, 2026 — the first major Five Eyes technical-and-policy document specifically on agentic AI deployment (Source: insideaipolicy.com). See Agentic AI, CISA — Cybersecurity and Infrastructure Security Agency (AI Deployer).
House appropriators warned, in a document reported June 24, 2026, that the Defense Department lacks visibility into the true compute costs of deploying AI across the enterprise, and directed the Pentagon to report regularly on its oversight of those costs (Source: insideaipolicy.com). The House Appropriations Committee subsequently attached report language to its passed defense spending bill seeking Pentagon details on autonomous-weapons "law of war" policies, AI infrastructure costs, and cybersecurity for the department's own AI systems, per accounts published July 3, 2026 (Source: insideaipolicy.com). The concern parallels the broader federal question — surfaced separately by the FERC and DOE data-center actions — that AI compute costs are poorly tracked even as deployment scales, and it adds a fiscal-oversight dimension to the department's rapid generative-AI adoption noted by vendors such as Shield AI.
At a June 30, 2026 gathering, senior Defense Department officials and industry executives said the U.S. military's next competitive advantage may depend less on designing the most sophisticated systems than on building them faster and at greater scale using digital engineering and artificial intelligence (Source: insideaipolicy.com).
On July 13, 2026, the House Armed Services Committee directed the Pentagon to formalize procedures for blacklisting AI companies from Defense Department systems beyond the Chinese firms already barred, and separately pressed it to examine AI-enabled manufacturing as a response to missile-production bottlenecks (Source: insideaipolicy.com; insideaipolicy.com). See FY27 National Defense Authorization Act (Chairman's Mark).
Representative Sara Jacobs (D-Calif.) introduced the Defense AI Reliability and Reporting Act on August 31, 2026, which would create an AI incident and vulnerability reporting programme within the Pentagon and a repository of AI weaknesses spanning procurement, testing, development and operation, as well as risks arising from human-machine teaming. Representatives George Whitesides (D-Calif.) and Nathaniel Moran (R-Texas) cosponsored the measure (Source: nextgov.com). It would give the department a standing internal counterpart to the incident-reporting duties surveyed at AI Incident Reporting, which reach developers rather than deploying agencies, and overlaps with the incident-and-vulnerability reporting requirements already carried in the FY27 NDAA Chairman's mark.
Proposed AI mandates have also reached parts of the department outside the warfighting and procurement debates. The Military Readiness Through Resilient Lands Act (H.R. 10069), introduced August 6, 2026 by Representatives George Whitesides, Chrissy Houlahan and Mike Haridopolos, would amend the Sikes Act to direct the Defense Department to assess and, where appropriate, apply AI to natural-resource and conservation planning on military installations (Source: nextgov.com). The bill is an instance of AI-assessment language attaching to an existing environmental statute rather than to defense-AI authorities.
Relationships
- deployed-by: Anthropic, Palantir Technologies, Anduril Industries, OpenAI (via ICE/federal channels).
- regulated-by: DoD Directive 3000.09 — Autonomy in Weapon Systems (source summary) (autonomous weapons directive), congressional oversight.
- depends-on: Chief Digital and AI Office (CDAO), Defense Innovation Unit (DIU), Radha Plumb.
- litigates: Anthropic v. United States (Pentagon ban challenge) — respondent in the D.C. Circuit petitions (Anthropic PBC v. U.S. Department of War — D.C. Circuit Opinion, Nos. 26-1049 and 26-1162 (September 25, 2026)); exclusion authority under Federal Acquisition Supply Chain Security Act of 2018 (FASCSA).
- related: Sens. Warner, Coons and Reed — letter to Secretary Hegseth and DNI Clayton on AI targeting and intelligence errors (2026), House Democrats' letter to Secretary Hegseth on civilian harm in Operation Epic Fury (2026) — congressional letters on AI targeting errors.
- related: Anthropic Claude Gov + Pentagon Dispute (2025–2026), Anthropic v. United States (Pentagon ban challenge), Autonomous Weapons, ICRC Position on Autonomous Weapon Systems (source summary), Pete Hegseth, Clawed, Federal AI Adoption — Patterns and Tensions.