The Black Box 013: The Pentagon's AI Chief Credited Grok With Helping Send Over 2,000 Munitions to 2,000 Targets in 96 Hours. His Office Told GAO Its Responsible AI Team Had 3 of 10 Positions Filled.

Bot Mutiny | October 23, 2026

A GAO audit released Oct. 14 describes the state of the Pentagon's Responsible AI program. Its text, as published, does not mention Grok.

The Black Box 013: The Pentagon's AI Chief Credited Grok With Helping Send Over 2,000 Munitions to 2,000 Targets in 96 Hours. His Office Told GAO Its Responsible AI Team Had 3 of 10 Positions Filled.

A GAO audit released Oct. 14 describes the state of the Pentagon's Responsible AI program. Its text, as published, does not mention Grok.

October 23, 2026

Footnote 27 of a Government Accountability Office report released on October 14 gives examples of how the Pentagon measures its Responsible AI tools.

It counts unique visitors to the toolkit's web page, how long they stay and actions such as downloads.

Those are examples that Responsible AI officials in the Chief Digital and Artificial Intelligence Office gave GAO of metrics they developed to measure people's experience with the tools. The same footnote says that of the 10 Defense Department components GAO surveyed, two used the office's Responsible AI tools on the AI project GAO asked them about. Seven did not. One did not know.

Four months before that report came out, the head of that office signed a declaration about Grok.

On June 15, 2026, the Justice Department asked a federal court in Mississippi to let it intervene in, and dismiss, a Clean Air Act lawsuit that the NAACP had filed against Elon Musk's xAI over gas turbines at a facility in Southaven, Mississippi, that power its data centers. Attached to the motion was a declaration from Cameron Stanley, the Department of War's Chief Digital and Artificial Intelligence Officer. The government's brief, citing it, says that during Operation Epic Fury, Grok, through the Pentagon's Maven Smart Systems, "enabled U.S. forces to deploy over 2,000 munitions to 2,000 distinct targets within 96 hours."

As of February 2026, according to CDAO officials cited in the GAO report, three of the 10 positions authorized for Responsible AI in the office were filled.

What was actually built

The Pentagon's Responsible AI program is older than the current administration and was built in layers. GAO's report, Artificial Intelligence: Actions and Opportunities to Implement Responsible Use in DOD (GAO-27-107837), lays them out.

In February 2020, the Secretary of Defense adopted the department's AI Ethical Principles. In May 2021, the Deputy Secretary reaffirmed them, defined six tenets for putting them into practice, and gave direction to the Responsible AI Working Council, which GAO describes as the department's permanent working-level collaborative group for the program. In June 2022, the department published the Responsible AI Strategy and Implementation Pathway, which set out 64 lines of effort and assigned them to offices across the department.

The first line of effort in the Pathway, according to GAO, was to fully staff the CDAO's Responsible AI office with expertise in AI technology, policy, acquisition, workforce and governance. CDAO officials told GAO that a 2023 resourcing memorandum authorized 10 positions, including a systems engineer, a program manager and an AI ethicist.

The design is federated: the CDAO writes guidance and builds tools, and the military services and defense agencies implement. The Working Council is where the two sides are supposed to meet, report progress and resolve problems. The Pathway also directed the CDAO to develop metrics measuring Responsible AI adoption.

The council. CDAO officials acknowledged it did not meet from late 2024 until August 2025. Responsible AI Division officials told GAO that other work, including a generative AI toolkit and a government-wide risk management document, kept them from organizing meetings. The council resumed in August 2025. It has no charter setting roles, responsibilities or meeting frequency, although the Pathway said it should be formalized through one. GAO wrote that without a charter, the council is at risk of lapsing again.

The lines of effort. As of April 2026, CDAO documentation showed 29 of the 64 were incomplete. That is 45 percent, nearly four years after the Pathway was published.

The metrics. GAO found the CDAO has tracked the lines of effort but has not developed metrics to measure implementation of Responsible AI. The tool metrics officials described to GAO, by their own examples, track use of the toolkit web page.

The staff. Three of 10 positions filled as of February 2026. CDAO officials told GAO the office is unlikely to fill the vacant positions or increase staffing, because it is still restructuring after an August 2025 move that placed it under the Under Secretary of Defense for Research and Engineering, and must account for the new 2026 strategy. GAO notes that the Congressional Research Service described that move as an effective demotion.

The definition. In January 2026, the Secretary of Defense issued the Artificial Intelligence Strategy for the Department of War. GAO summarizes it as emphasizing a wartime footing, calling for an "AI first" force, and clarifying that Responsible AI is to use models free from usage policy constraints that may limit lawful military applications, with model objectivity as a primary procurement criterion. Responsible AI Division officials told GAO they have been charged with aligning all existing AI policy guidance with that memo.

DOD concurred with both of GAO's recommendations, a charter for the council and staffing for the division, in comments GAO says it provided by email.

The innocent reading

The strongest case that none of this matters goes as follows.

Responsible AI is not the Pentagon's only control on targeting. Weapons and targeting decisions sit inside law-of-war review, rules of engagement and command authority, none of which GAO's report addresses. A three-person central team was never meant to review every use case; the Pathway assigns execution to the services and agencies, and GAO found most of the 10 components it surveyed had taken at least initial steps to set up oversight. Nine of 10 had established or begun establishing oversight bodies in their own guidance.

The tools also reached the system Stanley described. GAO reports that in December 2025 the Responsible AI Division integrated its toolkit into the Maven Smart System. The portions of the declaration quoted by the government and by Military Times describe Grok enabling forces to deploy munitions and supporting missions including targeting. They do not say Grok selected targets or that no human approved them.

And the staffing gap has an ordinary explanation: a reorganization and a new strategy that directed the office to align existing AI policy with it.

All of that may be true. The documents do not show that any Grok-assisted strike was wrong, and this piece does not claim it.

What the record leaves open

This piece asks a narrower question: whether the department's Responsible AI program, as GAO describes it, could show that it had assessed a model the Pentagon says it used at that scale. Six features of GAO's record bear on that.

No measures. GAO says the CDAO has no metrics for Responsible AI implementation.
The audit's sample was not drawn to find it. GAO drew its use cases from a January 2025 CDAO inventory of 378 use cases, fielded its survey from June 12 to August 14, 2025, and confirmed the information as of March 2026. Neither "Grok" nor "xAI" appears in the report's text. No document reviewed here says when Grok Gov entered Maven, so this piece does not claim the sample predates it.
The tools were rarely used on the projects GAO examined. Two of 10 surveyed components used the CDAO's Responsible AI tools for their selected project, per footnote 27, though CDAO documentation lists four organizations that have used the toolkit.
Proprietary models resist review. In GAO's Table 2, the Air Force reported that industry partners' proprietary AI models lacked transparency in data, algorithms or model characteristics, which hindered an objective assessment of Responsible AI implementation. GAO notes the Table 2 examples are illustrative and do not represent trends across DOD, and the Air Force comment concerned its own use case. Grok Gov is an industry partner's proprietary model.
Classified environments lack the tools. The Marine Corps told GAO that Responsible AI tools were not available in the classified environment. Six of 10 components said putting the CDAO's tools on classified domains would help. Military Times reported that, according to Stanley, Maven Smart System users process nearly two billion tokens a day on the department's top-secret network. GAO does not say on which network the December 2025 Maven integration of the toolkit runs.
The definition moved. The January 2026 strategy, as GAO describes it, prioritizes speed and clarified that Responsible AI is to use models free from usage policy constraints that may limit lawful military applications. Responsible AI Division officials told GAO they have been charged with aligning all existing AI policy guidance with it.

None of those six points depends on Grok. All six rest on GAO's text; points 4 and 5 add the government's brief and Military Times.

The named parties

Cameron Stanley, the Chief Digital and Artificial Intelligence Officer, signed the June 15 declaration. According to the Justice Department's brief, the declaration describes Grok as one of only four proprietary state-of-the-art AI models able to support national security applications. Military Times reported that he described Grok as deployed in Maven "to support vital national security missions, including targeting, intelligence, readiness, and recruitment." His office houses the Responsible AI Division whose work GAO reviewed.

The Justice Department, in a filing signed by Associate Attorney General Stanley E. Woodward Jr., used the declaration to argue that the NAACP's suit threatens national security, and asked the court to dismiss it outright on the ground that the Clean Air Act's intervention right, 42 U.S.C. § 7604(c)(2), carries a right to end citizen suits. The United States then appealed to the Fifth Circuit what the xAI brief calls the district court's constructive denial of intervention, and now seeks a stay.

xAI (now SpaceXAI Corp.) and MZX Tech, represented by Paul D. Clement among others, filed an amicus brief in the Fifth Circuit, dated October 12, supporting a stay. It states that "the American military relies on Grok for targeting operations in an ongoing military conflict," citing the Stanley declaration and a declaration by Jacobson (Dkt. 94-3). The brief argues a shutdown would leave the Grok tool effectively unable to function and cost the companies billions of dollars.

GAO's Defense Capabilities and Management team, under Director Alissa H. Czyz, conducted the audit from September 2024 to October 2026 under a provision in Senate Report 118-188 accompanying the fiscal 2025 defense authorization bill. Its recommendations went to the Secretary of Defense. It also asked Congress to consider directing DOD to implement open recommendations from 2022 and 2023 on AI collaboration and the AI workforce, which GAO says remain open or partially addressed.

The NAACP, represented by Earthjustice and the Southern Environmental Law Center, alleges that xAI ran 27 turbines at the Southaven plant without required permits. Mississippi regulators determined no permit was needed. That dispute is about air permits, not targeting, and it is unresolved.

What the instrument cannot see

The Responsible AI program was designed to make AI systems trustworthy enough to speed their adoption, through assessment, test and evaluation, and monitoring. The Pathway routes that work through component leads, a council and a central office.

GAO's report shows the route as of 2026. The central office had three of 10 authorized positions filled, and officials said it is unlikely to fill the rest. The council did not meet from late 2024 until August 2025 and has no charter. The tool metrics officials gave as examples track toolkit page visitors, visit duration and downloads. One component told GAO the tools were unavailable in its classified environment, though the toolkit was integrated into the Maven Smart System in December 2025. And the newest strategy, as GAO summarizes it, says Responsible AI is to use models free from usage policy constraints that may limit lawful military applications.

What none of those documents can tell a reader is whether Grok Gov went through a Responsible AI review before it reached Maven. The GAO report does not address it. The declaration, as quoted in the government's brief and by Military Times, does not address it. That absence is a gap in the record, not evidence that no review happened.

Black Box 010 covered the UN report on the February 28 strike on a school in Minab, Iran. That report does not mention artificial intelligence, and nothing in the documents reviewed here connects Grok, Maven or any AI system to that strike. This piece does not either.

The sequence

DOD concurred with both recommendations, in comments GAO says it provided by email. Earlier, CDAO officials had told GAO they were unlikely to fill the vacant positions. The report does not say whether the concurrence changes that.

The order of events matters. The council stopped meeting while the Responsible AI Division built a generative AI toolkit. In December 2025, the division put its toolkit into Maven. In January 2026, the strategy clarified the definition and scope of Responsible AI. As of February, three of 10 positions were filled. In June, the office's chief filed a declaration that the government cites for the claim that Grok, through Maven, enabled more than 2,000 munitions to be deployed in 96 hours, and the government used that figure in court to argue that a pollution lawsuit should be dismissed.

The government's brief, citing the CDAO's declaration, told a federal court that Grok's continued operation and availability "is a matter of paramount national security." Nothing in the GAO report or the filings reviewed here says how, or whether, Grok was checked under the Responsible AI program.

The record

What the documents establish

GAO-27-107837, released Oct. 14, 2026, reports that CDAO officials said only three of 10 authorized Responsible AI positions were filled as of February 2026, and that the office was unlikely to fill the rest.
The Responsible AI Working Council did not meet from late 2024 until August 2025 and has no charter.
As of April 2026, 29 of 64 Pathway lines of effort were incomplete, and the CDAO had no metrics for Responsible AI implementation.
Examples of the Responsible AI tool metrics CDAO described to GAO include web-page visitors, visit duration and actions such as downloads.
Two of 10 surveyed components used CDAO's Responsible AI tools on their selected project.
The Responsible AI toolkit was integrated into the Maven Smart System in December 2025.
The GAO report's text does not mention Grok or xAI. Its survey sample came from a January 2025 inventory.
In a June 15, 2026 filing, the Justice Department, citing CDAO Cameron Stanley's declaration, said that during Operation Epic Fury Grok, through Maven, "enabled U.S. forces to deploy over 2,000 munitions to 2,000 distinct targets within 96 hours."
xAI's Oct. 12 Fifth Circuit brief states the military relies on Grok for targeting operations in an ongoing conflict.
DOD concurred with GAO's two recommendations, by email.

What the documents do not establish

Whether Grok Gov went through any Responsible AI review, law-of-war review or other assessment before or during Operation Epic Fury.
When Grok Gov entered the Maven Smart System.
What Grok did in the targeting process: whether it nominated targets, summarized intelligence, managed logistics or something else.
Whether humans approved each of the 2,000 targets.
Whether any Grok-assisted strike was inaccurate or unlawful.
Any connection between Grok, Maven or AI and the Minab school strike.
Whether the Responsible AI staffing changed between February and October 2026.
Whether the turbines at Southaven violated the Clean Air Act. That question is before the courts.

Close

The Pentagon set up its Responsible AI program to build trust in its AI through assessment, testing and monitoring. In June 2026 the government cited its AI chief's declaration, which credits one model with enabling more than 2,000 munitions in 96 hours, to seek dismissal of a pollution suit. Neither GAO's report nor the filings reviewed here say whether that model went through the program.

Sources: U.S. Government Accountability Office, Artificial Intelligence: Actions and Opportunities to Implement Responsible Use in DOD, GAO-27-107837 (Oct. 14, 2026), including footnotes 11 and 27, Table 2, Table 3 and Appendix III. The United States' Motion for Intervention and Dismissal and Memorandum in Support, NAACP v. X.AI, No. 3:26-cv-00074-DMB-JMV (N.D. Miss.), Docs. 58 and 59 (filed June 15, 2026), as posted by the Department of Justice; Exhibit A (Doc. 58-1), the declaration of Cameron Stanley, is image-only in that posting and is cited here as quoted in the government's memorandum and by Military Times. Amicus brief of X.AI Corp. (now SpaceXAI Corp.) and MZX Tech LLC supporting the motion for stay pending appeal, NAACP v. X.AI Corp., No. 26-60806 (5th Cir., dated Oct. 12, 2026). 42 U.S.C. § 7604. Daniel Terrill, "The unlikely role of Operation Epic Fury in a Mississippi AI data center lawsuit," Military Times, June 22, 2026. UN Human Rights Council, A/HRC/63/61, advance edited version (Sept. 21, 2026). Earlier Bot Mutiny coverage:botmutiny.com/blog/xai-tells-fifth-circuit-the-military-relies-on-grok-for-targeting-in-bid-to-pause-naacp-uhbo78t3,botmutiny.com/blog/pentagon-responsible-ai-gao-says-council-did-not-meet-for-months-and-only-3-of-10-cjxrdlnrandbotmutiny.com/blog/the-black-box-010-minab-un-report-never-mentions-ai.