Anthropic Holds the Line as Pentagon Pushes AI Safeguard Removal

Anthropic Stands Alone Against Pentagon Pressure to Strip Military AI Safeguards

Daily Signal — March 8, 2026

TL;DR: Secretary of Defense Pete Hegseth has given Anthropic an ultimatum — remove restrictions on mass domestic surveillance and autonomous weapons use or face designation as a “supply chain risk” or forced compliance under the Defense Production Act. Anthropic refused; OpenAI, Google, and xAI have already accepted the Pentagon’s “all lawful purposes” standard. Congress is now moving to constrain the executive branch’s ability to use coercive procurement tools against AI companies, while public polling shows Americans broadly support the safeguard positions Anthropic is defending.

Today’s Themes

Whether private AI companies retain any principled right to impose usage constraints on government customers — or whether federal procurement power effectively nullifies that right.
The Pentagon’s “all lawful purposes” framing versus the existing Defense Directive 3000.09 requirement for “meaningful human control” over lethal force — two stated U.S. government positions now in direct contradiction.
The Defense Production Act as a coercive instrument: whether its threatened application against a domestic company over a contract dispute represents a legitimate or abusive use of emergency authority.
The emerging split among major AI labs between companies that treat safety constraints as negotiable and one that, so far, does not.
Whether democratic legitimacy arguments — “elected governments should decide, not private companies” — hold when the specific decisions at issue involve warrantless surveillance and autonomous lethal force.

Top Stories

Pentagon Pressures AI Companies to Remove Safeguards on Military Contracts

What happened: Secretary of Defense Pete Hegseth has demanded that AI companies operating under Pentagon contracts remove usage restrictions and accept a new “all lawful purposes” policy. Anthropic has specifically refused to lift prohibitions on mass domestic surveillance and fully autonomous weapons. OpenAI, Google, and xAI have agreed to the Pentagon’s terms. Hegseth set a February 28 deadline for Anthropic’s compliance and threatened to designate the company a “supply chain risk” or invoke the Defense Production Act to compel it. Common Cause and allied organizations have formally urged Congressional oversight of the dispute. The contract at the center of the dispute is valued at $200 million.

Why it matters: Builders, operators, and investors across the defense AI sector now face a concrete precedent-setting event, not a hypothetical. The mechanism here is procurement leverage: the Pentagon is using contract access and emergency statutory authority to restructure a private company’s internal product policy. If this succeeds, the effective message to every AI lab seeking federal revenue is that safety constraints are a negotiating position to be surrendered, not an engineering or ethical baseline. The contradiction with Defense Directive 3000.09 — which already requires “meaningful human control” over lethal force — is not incidental. It signals that the current administration is willing to operate outside its own standing directives, which means legal exposure for any company that complies and later faces liability for harms caused by systems that violated those directives. Compliance is not a safe harbor.

Contract value: $200 million
Deadline issued to Anthropic: February 28, 2026
Companies that accepted Pentagon terms: OpenAI, Google, xAI
Company refusing: Anthropic (sole holdout among named major labs)
Threatened enforcement mechanisms: “supply chain risk” designation; Defense Production Act invocation
Existing counteracting directive: Defense Directive 3000.09 — requires “meaningful human control” over lethal force

Sources: commoncause.org, anthropic.com

Congress Prepares Legislative Response to Pentagon AI Dispute

What happened: Representative Sam Liccardo (D-Calif.) plans to introduce an amendment to the Defense Production Act that would prohibit federal agencies from retaliating against AI and emerging technology providers in contract disputes. Senate Democrats are in early discussions about broader legislation specifically addressing Pentagon pursuit of AI for domestic surveillance and autonomous weapons. A bipartisan group of Armed Services Committee leaders sent a private letter urging settlement between Anthropic and the Pentagon. The Defense Production Act is scheduled to expire September 30, 2026.

Why it matters: The DPA’s September 30 expiration date is the operative lever here. Liccardo’s amendment is structurally timed to attach to reauthorization negotiations, which gives it more procedural traction than a standalone bill would have. For AI companies currently evaluating or holding defense contracts, this legislative window is the most plausible near-term structural protection — but it is not yet law, it is bipartisan only at the letter-writing level, and it faces a Senate environment where broader legislation is explicitly described as “early discussions.” Companies should not treat Congressional momentum as resolved protection. The bipartisan Armed Services letter urging settlement rather than legislation also signals that some members would prefer this dispute not become statute-generating — which means outcomes could be negotiated quietly before the legislative calendar matures.

Lead sponsor: Rep. Sam Liccardo (D-Calif.)
Target legislation: Amendment to the Defense Production Act
DPA expiration date: September 30, 2026
Bipartisan Armed Services letter: urges settlement, not legislation
Senate action status: early discussions only, as of reporting date

Source: axios.com

Public Debate Intensifies Over AI Companies’ Role in Military and Surveillance

What happened: OpenAI CEO Sam Altman faced direct public questions about his company’s willingness to support mass surveillance and autonomous killing — the specific activities Anthropic declined to enable. Altman stated that private companies should not set national policy and that such decisions belong to civilian democratic leadership. He characterized his level of surprise at the intensity of public debate as notable.

Why it matters: Altman’s deference-to-democracy framing is not a neutral position — it is a substantive argument that has meaningful implications for how responsibility is allocated if OpenAI’s systems are used for activities that cause legal or civilian harm. Framing compliance as “deferring to elected leadership” transfers moral and, potentially, legal accountability entirely to government actors. This matters for OpenAI’s institutional posture and for the broader sector, because it normalizes a model in which AI companies disclaim product-level responsibility for end uses sanctioned by government customers. Whether courts and regulators will accept that framing remains an open question, particularly given the Fourth Amendment dimensions raised in the public polling data.

OpenAI has accepted Pentagon “all lawful purposes” terms
Altman’s stated position: private companies should defer to civilian democratic leadership on national policy
Altman acknowledged surprise at the scale of public debate over the question

Source: techcrunch.com

Americans Overwhelmingly Support Human Control in Military AI Decisions

What happened: A survey published by ITIF found that 79% of Americans believe a human being should always make the final decision before lethal force is used, with 81% agreement among both Democrats and Republicans. Separately, 75% said AI is not yet reliable enough for life-or-death military decisions without human oversight; 70% said warrantless AI surveillance of Americans violates the Fourth Amendment; 67% said technology companies have a responsibility to set limits on product use; and 53% supported AI companies restricting their technology from domestic surveillance or autonomous weapons applications.

Why it matters: For Congressional offices currently evaluating the Liccardo amendment and broader Senate legislation, this polling data provides unusually direct political cover — and it is bipartisan cover, which matters for vote-counting arithmetic in a closely divided legislature. The 81/81 partisan split on human control of lethal force is the most significant data point: it means this is not a wedge issue the administration can use to peel off Republican votes against oversight legislation. The 70% figure on Fourth Amendment concerns is separately important because it frames the domestic surveillance dimension of the Pentagon’s “all lawful purposes” policy as constitutionally suspect in the public mind — which increases litigation risk for any administration program built on unrestricted AI surveillance capabilities.

79% of Americans: humans must always make final lethal force decisions
81% Democrats, 81% Republicans: same position, indicating no partisan gap
75%: AI not reliable enough for military life-or-death decisions without oversight
70%: warrantless AI surveillance of Americans violates the Fourth Amendment
67%: tech companies have a responsibility to set limits on product use
53%: support AI companies restricting technology from domestic surveillance or autonomous weapons
Source organization: ITIF (Information Technology and Innovation Foundation)
Publication date: February 26, 2026

Source: itif.org

Security Watch

DPA as coercive tool against domestic companies: The Pentagon’s threatened invocation of emergency procurement authority against Anthropic — a U.S. company raising constitutional objections — has no established legal precedent in the AI context. The mechanism and its limits are untested.
Mass domestic surveillance without legal framework: The “all lawful purposes” policy, if implemented with AI systems at scale, would precede any statutory or judicial framework governing its use. The gap between capability deployment and legal authorization represents a significant institutional risk.
Autonomous weapons without meaningful human control: The Pentagon’s push to remove safeguards sits in direct tension with Defense Directive 3000.09, creating an internal legal inconsistency within DOD that has not been publicly resolved.
“Supply chain risk” designation applied to American AI companies: The designation is typically associated with foreign adversary suppliers. Its threatened application to a U.S. company in a domestic contract dispute is a novel and potentially precedent-setting use of the classification.

What to Watch Next

Anthropic’s post-deadline posture: The February 28 compliance deadline has passed as of this briefing’s publication date. Whether Anthropic has formally refused, sought extension, or entered negotiation — and whether DOD has responded with any enforcement action — is the immediate question requiring confirmation.
DPA reauthorization timeline in Congress: Watch for whether Liccardo’s amendment is formally introduced and whether it is attached to the DPA reauthorization markup. Committee scheduling before the September 30 expiration will determine whether the legislative window is viable or symbolic.
Senate Democrats’ legislative language: “Early discussions” on broader surveillance and autonomous weapons legislation will either mature into a draft bill or stall. Senate Armed Services Committee membership and any movement from the Republican side of the bipartisan letter are the indicators to track.
DOD reconciliation of “all lawful purposes” with Directive 3000.09: The administration has not publicly explained how removing meaningful human control requirements from AI contracts is consistent with its own standing lethal-force directive. Any formal DOD statement on this contradiction — or a quiet revision of the directive — would be a significant development.
OpenAI, Google, and xAI contract implementation: Now that three major labs have accepted Pentagon terms, the specific applications and oversight structures governing those contracts will begin to take shape. Any public disclosure, congressional inquiry, or FOIA-responsive documentation on implementation scope would materially inform the risk picture.

Sources

AI-generated editorial illustration · TemperatureZero · March 8, 2026

Keep reading the signal

Get the Daily Signal — a concise briefing on what actually matters in AI and the systems around it.

Subscribe Free