The Black Box 002: DHS Agreed With Every AI Warning Its Watchdog Issued. Then It Scaled Up Anyway.
Bot Mutiny |
20 findings, a 40 percent AI surge, a $1B Palantir ceiling, and a viral claim that does not survive the registry
The Black Box 002: DHS Agreed With Every AI Warning Its Watchdog Issued. Then It Scaled Up Anyway.
20 findings, a 40 percent AI surge, a $1B Palantir ceiling, and a viral claim that does not survive the registry
July 17, 2026
Gold Without the Games. Faith-Driven. Fully Transparent.
Most gold companies profit from your confusion. True Gold Republic was built differently. America's only Christian precious metals firm, with transparent pricing, zero hidden fees, and a written guarantee. Get your free 2026 kit and see the difference.
Keep up with tech in 5 minutes. TLDR is the free daily email of the most interesting stories in startups, tech, and programming, curated by ex-Google and Anthropic engineers.
Ignore the press release. Read the documents.
The paper trail
In January 2025, the Department of Homeland Security's own Inspector General put a warning in writing. The report, OIG-25-10, issued January 30, 2025, credited the department for issuing AI-specific guidance, appointing a Chief AI Officer, and standing up multiple working groups and an AI Task Force to steer its artificial intelligence efforts. Then it made 20 formal recommendations to improve DHS governance and oversight of AI. The report's core sentence was not subtle. In the Inspector General's own words, without appropriate, ongoing governance, DHS faces an increased risk that its AI efforts will infringe upon the safety and rights of the American people.
That sentence is not a paraphrase pulled from secondary coverage. It appears in the OIG's Semiannual Report to Congress covering October 2024 through March 2025, the office's formal accounting to lawmakers, which summarizes OIG-25-10 and confirms the report's structure: 20 recommendations, and a department that concurred with all 20.
DHS did not dispute a single finding. The concurrence is recorded in the OIG's own summary and was reported contemporaneously by Nextgov/FCW and HSToday.
That was the promise. This investigation is about what the public record shows DHS did in the eighteen months that followed. The short version: the department's AI footprint grew faster than at any point in its history, a billion dollar ceiling contract with Palantir went live across its major components, and the compliance paperwork that federal law requires as the public's only window into these systems stayed partial and contested. Along the way we will correct two viral claims about that record, because the truth is damning enough without exaggeration, and because Bot Mutiny does not launder numbers it cannot trace to a primary document.
What DHS agreed to
The Inspector General's semiannual report walks through what the audit actually found, and the findings sort into three failures that matter more together than they do apart.
The first is strategy without execution. DHS established an AI strategy to guide enterprise-wide goals and objectives, but the OIG found the department did not effectively execute that strategy because it never developed an implementation plan. A strategy with no implementation plan is a press release with a longer shelf life.
The second is oversight without capacity. The OIG concluded that DHS lacked adequate governance processes to monitor whether its AI use complied with privacy and civil liberties requirements, and attributed the gap to resource challenges. The DHS Privacy Office answered that it completed a staffing assessment on January 31, 2025, one day after the report published, and determined its Privacy Oversight Team had the necessary resources in place to handle AI oversight. Whether a one-day turnaround reflects a rigorous assessment or a fast one is a judgment the reader can make.
The third failure is the one that should reframe every number DHS publishes about its own AI use, including the growth figure this piece is built around. The OIG found that DHS developed processes to track and report its AI use to the public, as required, but those processes did not identify and track some of the data needed to report the department's use cases. More pointed still: DHS had limited evidence to demonstrate why it considered its AI use consistent with federal requirements, because neither the department nor its components had a formalized process to identify, review, and validate the data included in DHS's mandated AI reporting. Hold that finding. It returns at the end of this piece.
None of this arrived in a vacuum. A month before OIG-25-10, in December 2024, the same office published OIG-25-06 on the department's early efforts to use AI for intelligence collection and analysis, with eight recommendations of its own. And the Government Accountability Office had already flagged, in a February 2024 report on DHS's use of AI for cybersecurity, that the department's public AI inventory contained entries that were not verified as actually being AI, including one of the two cybersecurity use cases GAO examined. DHS concurred there too.
So where do the 20 recommendations stand a year and a half later? The OIG maintains a public register of unresolved open recommendations for exactly this question, and as of the most recent published register, dated September 30, 2025, OIG-25-10 does not appear on it. In the vocabulary of federal oversight, that means all 20 recommendations are resolved: DHS and its watchdog have agreed on a corrective action plan for each one. Resolved is not the same as done. The OIG's own reporting distinguishes resolved recommendations from closed ones, and closure requires the department to actually complete the corrective work. Whether DHS has implemented any of the 20 fixes is not visible in the public record. What the record shows is a department that agreed to a to-do list, and a public with no way to check whether the list got done.
What DHS did instead
On January 28, 2026, DHS published its updated AI use case inventory. Analysis of that inventory by TechPolicy.Press found more than 200 active AI use cases across the department, an increase of nearly 40 percent since July 2025. Per the same analysis, Immigration and Customs Enforcement added 24 new applications in that window, including tools built on Palantir platforms.
The department's September 2025 compliance plan for OMB Memorandum M-25-21 tells the same growth story in its own words. The document describes AI innovation sandboxes, a DHSChat assistant deployed to all personnel, real-time AI risk assessment of passengers and vehicles at ports of entry, and a governance board that satisfied its convening requirement with a paper-only meeting dated July 2, 2025. A governance board that meets on paper is a governance board that does not deliberate.
Then came the money.
The Palantir ceiling, by the numbers
Federal procurement records settle what months of commentary blurred. USAspending lists the agreement under PIID 70RTAC26A00000001: a single-award blanket purchase agreement issued by the DHS Office of Procurement Operations, with performance beginning February 12, 2026, a potential value of one billion dollars, and placement against Palantir's GSA schedule contract 47QTCA24D004L.
The claim that this was a no-bid deal is wrong, and the government's own coding says so. The award's extent-competed field reads "A - Full and Open Competition." That code carries a caveat worth stating plainly: it reflects competition at the schedule level, meaning Palantir competed to get onto the GSA schedule in the first place, not a fresh head-to-head competition for this specific billion dollar vehicle. The distinction cuts both ways. Critics who call the BPA no-bid are wrong on the record. Anyone waving the competition code as proof this ceiling was freshly competed is stretching the record just as far.
The historical contrast makes the point sharper. Palantir's earlier ICE contracts for the FALCON analytics platform, awarded in 2015 and 2018, carried the extent-competed code "C - Not Competed." The paper trail has gotten cleaner. What flows under it has not.
In the BPA's first five months, three call orders totaling 150.8 million dollars appear in the record. A Secret Service data modeling order for 18.7 million dollars, dated March 18. An ICE Enforcement and Removal Operations modernization order for 86.3 million dollars, dated May 5. And an ICE Homeland Security Investigations order for investigative case management and analytics, 45.8 million dollars, dated June 26. Over 132 million of the 150.8 million, nearly nine of every ten dollars obligated so far, went to ICE enforcement systems. A single-award structure means that for the life of the agreement, one vendor sits inside the fence while competitors stay outside it, and each of those task orders received nothing like the scrutiny of a fresh open competition.
The access fight
While the AI estate grew, the department's relationship with its own watchdog deteriorated in a way you can read in the watchdog's formal filings. A claim circulated online that the Inspector General had accused DHS of obstructing audits of its biometric data management. That claim gets both the verb and the subject wrong, and the OIG's actual account is worse in its specifics than the paraphrase ever was.
The Inspector General's Semiannual Report to Congress for October 2024 through March 2025 contains a section the IG Act requires it to publish: a summary of incidents where the department restricted or significantly delayed the OIG's access to information. During that period, three DHS components, CBP, FEMA, and the Secret Service, denied and delayed OIG requests for direct read-only access to their databases or equivalent data extracts. The OIG states that department representatives did not cite any legal authority consistent with section 406(a)(1)(B) of the IG Act to justify withholding or delaying access. Instead, the components cited an inability to partition data, audit scope, and data protection concerns.
The specifics come from the OIG's own tables. FEMA denied the OIG direct read-only access to its FEMA GO grants database outright, then took 114 calendar days to deliver a complete set of Shelter and Services Program grant data, and 63 calendar days to produce records of its oversight of 425 million dollars in humanitarian subawards. Both delays, in the OIG's words, significantly impacted its ability to meet project milestones.
CBP denied read-only access to its BorderStat application for an audit of passenger inspections, after two prior denials of the same access on earlier projects. The extracts CBP provided in place of access proved incomplete and unreliable: the OIG discovered that relevant inspection records were coded incorrectly in the system, leaving records out of the extracts entirely, and that the numbers CBP sent did not match totals provided by CBP's own field offices.
The Secret Service took 126 calendar days to provision access to its Protective Threat Management System, which the OIG needed for its reviews of the July 2024 Butler, Pennsylvania assassination attempt. As of March 31, 2025, the Secret Service still had not provided access to its Concur travel system 131 days after the request, or to its Sunflower property management database 110 days after the request.
None of this involves biometrics, and the OIG's verbs are denied, delayed, and restricted, not obstructed. Precision here is not pedantry. An agency stonewalling its watchdog on grant money and assassination-attempt records is a documented fact with day counts attached. An agency obstructing biometric audits is a claim nobody has produced a document for. Bot Mutiny will take the documented fact.
The paperwork that tells you what your government knows about you
The Privacy Impact Assessment is one of the few legally mandated windows the public has into federal data systems. Section 208 of the E-Government Act of 2002 requires agencies to conduct a PIA before developing or procuring IT systems that collect, maintain, or disseminate information in identifiable form about members of the public. The assessments state what personal information a system collects, why, and how it is used and shared. When an agency deploys surveillance-adjacent technology without a corresponding PIA, the public loses its primary means of knowing the system exists at all.
In May 2026, a claim began circulating that DHS had filed zero Privacy Impact Assessments in all of 2026, against 24 in 2024 and 8 in 2025. The claim originated with surveillance-tracking projects including Capture Cascade and State of Surveillance, and it spread because it is a clean, brutal number attached to a real trend people are right to worry about.
It is also wrong, and DHS's own registry proves it. Immigration and Customs Enforcement published a PIA covering the immigration bond process with a listed date of January 29, 2026. Customs and Border Protection published a PIA on its Global Interoperability Standards trade-data testing, listed June 1, 2026. Whatever methodology produced zero, it does not survive contact with the registry it claims to describe.
Why does Bot Mutiny spend a section of its own investigation correcting critics of the agency it is investigating? Two reasons. First, because a false number handed to DHS is a gift: the department gets to rebut the exaggeration and walk away from the underlying question. Second, because the underlying question survives the correction. Two published assessments in six and a half months is not a transparency success story for a department whose AI inventory grew 40 percent in six months while a billion dollar surveillance platform ceiling went live underneath it. And the department's own registry caveats that PIAs are withheld from publication entirely when the subject matter is law enforcement sensitive or involves a national security system. The public count is a floor, and the systems most likely to concern the public are the ones least likely to appear in it.
What this looks like from the outside
Set the documents side by side and the record tells its own story, no speculation about motive required. In January 2025, the department's watchdog issued 20 formal recommendations about AI governance and warned, in writing, of an increased risk to the safety and rights of the American people. The department agreed with every one. Over the following year, that same department grew its AI inventory by 40 percent, deployed an AI assistant to its entire workforce, put AI risk-scoring on travelers at ports of entry, and stood up a billion dollar single-award purchasing vehicle with Palantir under which nearly nine of every ten obligated dollars has gone to ICE enforcement systems.
In the same window, by the watchdog's own formal account to Congress, three of the department's components denied or delayed the watchdog's access to their data systems without citing any legal authority for doing so. The public's mandated paper trail produced two published privacy assessments in six and a half months, with an unknown number withheld as law enforcement sensitive. And the status of the 20 fixes the department agreed to make is resolved on paper and unverifiable in practice. Every institution that could check the department's homework has been slowed, shorted, or left waiting. That is what the documents show.
What we cannot see
Every claim above rests on public documents: the OIG's published reports and its semiannual account to Congress, the department's own AI inventory and compliance plan, the PIA registry, federal procurement records, and named press reporting. What the public record cannot show is the substance behind the numbers. The three Palantir call orders are public entries in a procurement database; what the systems they fund actually do is not. The PIA registry lists what was published, not what was withheld. And the 20 recommendations sit in the resolved column with no public evidence of implementation. Those are the black boxes.
THE NUMBER TO REMEMBER
40 percent in six months. That is how fast the AI inventory of the department that concurred with 20 watchdog findings about inadequate AI governance grew, per its own January 28, 2026 inventory, while a one billion dollar Palantir ceiling went live underneath it. And per that same watchdog, DHS has no formalized process to identify, review, and validate the data in its mandated AI reporting. The department scaled up on the strength of a number its own Inspector General says it cannot verify.
Sources
DHS Office of Inspector General, OIG-25-10, January 30, 2025. DHS Office of Inspector General, Semiannual Report to Congress, October 1, 2024 through March 31, 2025, including its Summary of Restricted or Delayed Access to Information. DHS Office of Inspector General OIG-25-06, December 16, 2024. DHS OIG Unresolved Open Recommendations register, September 30, 2025 edition, accessed July 16, 2026. GAO-24-106246, February 2024. DHS AI Use Case Inventory, updated January 28, 2026. DHS Compliance Plan for OMB M-25-21, September 2025. DHS Privacy Impact Assessments registry, accessed July 16, 2026. USAspending.gov record for PIID 70RTAC26A00000001 and associated call orders, accessed July 16, 2026. FedScoop reporting on the Palantir blanket purchase agreement, February 2026. Nextgov/FCW, February 3, 2025. TechPolicy.Press analysis of the January 2026 DHS AI inventory. Capture Cascade and State of Surveillance tracker entries, cited as the origin of the disputed zero-PIA and audit-obstruction claims, not as evidence.
The Black Box is Bot Mutiny's primary-document investigative series. Corrections: human@botmutiny.com.