— The Pentagon has summoned Anthropic’s chief executive, Dario Amodei, to the Pentagon for a meeting with Defense Secretary Pete Hegseth amid a standoff over how the company’s A.I. models can be used on classified systems. The visit follows a renegotiation of a $200 million pilot contract after a Jan. 9 memo from Secretary Hegseth pressed A.I. firms to lift certain restrictions. Pentagon officials say they want Anthropic to accept the same operational terms being discussed with other vendors, while Anthropic insists on specific safeguards to prevent domestic mass surveillance and fully autonomous weapons without human oversight. The meeting is intended to resolve whether Anthropic will loosen limits and, if so, what safety guarantees will accompany broader military use.
Key Takeaways
- Anthropic’s CEO Dario Amodei is meeting Defense Secretary Pete Hegseth at the Pentagon to discuss access to classified systems; the report is dated Feb. 23, 2026.
- The two sides previously agreed on a $200 million pilot contract; that agreement is being renegotiated following a Jan. 9 memo from Hegseth urging removal of model restrictions.
- The Pentagon has already reached an agreement with xAI and is near a deal with Google’s Gemini model, using those discussions as leverage in talks with Anthropic.
- Pentagon negotiators want contracts that allow the department to use models as it deems lawful while still permitting companies to embed safety features, the so-called “safety stack.”
- Anthropic was the first company authorized to operate on the military’s classified networks but has conditioned broader access on guardrails that bar mass domestic surveillance and weapons without humans in the loop.
- Google and xAI did not provide immediate comment to reporters about their contract terms or the Pentagon meeting.
Background
The Defense Department began accelerating ties with commercial A.I. firms after years of investment and experimentation with large language and multimodal models. Anthropic secured initial access to classified networks, positioning it as an early partner in military A.I. work. The company’s Claude model and safety-oriented design have made it a focal point in debates about how private-sector safety practices align with operational military needs.
Tensions escalated when Secretary Hegseth issued a Jan. 9 memo urging A.I. providers to eliminate restrictive guardrails that could limit Defense Department use. That directive triggered renegotiations of existing pilot agreements, including Anthropic’s $200 million pilot. Simultaneously, the Pentagon has pursued parallel pacts with other vendors — including an executed agreement with xAI and advanced talks with Google — creating a bargaining dynamic across providers.
Main Event
Pentagon officials summoned Dario Amodei to Washington to press for commitments aligning Anthropic’s contract with terms the department is seeking from other firms. Officials conveyed that the department seeks contractual authority to use models for lawful operations while allowing firms to retain technical safety measures. The meeting is being framed as a narrow negotiation over contractual language and operational scope rather than a broader public confrontation.
Anthropic’s negotiating position, according to people involved in the discussions, is that it will relax some internal limits only if explicit safeguards are written into any expanded agreement. Those safeguards, they say, must forbid the models’ use in large-scale domestic surveillance and block deployment in weapons systems that operate without a human decision-maker in the loop. Anthropic’s role as the first company cleared for classified systems gives it leverage but also places it under heightened scrutiny.
Pentagon officials have indicated they will ask Anthropic to accept the same guardrail framework being negotiated with xAI and Google. That framework, as described by negotiators, would permit the department operational flexibility while preserving a place for company-implemented safety features. Observers say the administration’s public pressure and alternative vendor agreements are intended to encourage convergence on terms favorable to the Defense Department.
Analysis & Implications
The clash reflects a broader policy tension: the Defense Department’s operational priorities versus commercial firms’ commitments to safety and public trust. For the Pentagon, unrestricted access to high-capability models can accelerate intelligence, logistics and decision-support capabilities; for firms like Anthropic, assurances about lawful use and reputational risk shape what they will accept. The outcome of these negotiations will set a precedent for future contracts and for how civilian A.I. developers balance safety and national-security demands.
If Anthropic accedes to the Pentagon’s terms, other companies may face similar pressure to harmonize contractual language, effectively narrowing the space for individualized safety policies. Conversely, if Anthropic secures enforceable prohibitions on mass domestic surveillance and fully autonomous weaponization, it could institutionalize stronger safety constraints across Defense Department procurements. That result would complicate rapid operational deployment but could reduce legal and ethical liabilities for vendors.
Internationally, U.S. precedent will matter: allied militaries and foreign governments watch how commercial A.I. firms and the Pentagon resolve these issues. A bargain that preserves some company-controlled safety mechanisms while granting the Defense Department lawful operational use could become a model for allied procurement. Alternatively, a shift toward government-mandated permissiveness could spur regulatory debates, corporate exit decisions, or alternative procurement strategies abroad.
Comparison & Data
| Company | Model | Contract Status | Notable Constraints |
|---|---|---|---|
| Anthropic | Claude | Renegotiating $200M pilot | Wants safeguards against mass domestic surveillance and fully autonomous weapons |
| xAI | xAI model | Signed agreement with Pentagon | Reportedly more permissive operational terms |
| Gemini | Close to a deal | Negotiations ongoing; safety provisions allowed |
The table summarizes publicly reported contract status and the core constraints discussed. While the Pentagon seeks a consistent operational baseline, companies differ on what technical or contractual safeguards they will accept. The $200 million figure for Anthropic’s pilot provides scale: these are nontrivial commercial engagements that will influence vendor risk calculations and the government’s procurement strategy.
Reactions & Quotes
“The secretary confirmed he will meet with Mr. Amodei at the Pentagon to discuss the contract terms.”
Defense Department official (confirmed to reporters)
Context: Pentagon spokespeople declined to preview announcements but acknowledged the scheduled meeting and framed it as part of ongoing negotiations over operational use and safety language.
“Anthropic says it will ease limits only if written safeguards prevent mass domestic surveillance and fully autonomous weapons without human oversight.”
People involved in the discussions (reported to press)
Context: Sources close to the talks described Anthropic’s red lines as focused on civil liberties and weapons policy, emphasizing legally enforceable language rather than purely technical controls.
Unconfirmed
- Whether the Pentagon will ultimately require identical contractual language from all vendors remains uncertain; negotiations with Google and xAI differ in details and are ongoing.
- The precise technical form of the “safety stack” that would satisfy both the Defense Department and Anthropic has not been published and may vary across deployments.
- No public statement has confirmed whether new language would categorically prohibit specific future uses beyond the examples discussed (mass surveillance, autonomous weapons).
Bottom Line
This meeting pits a leading safety-focused A.I. company against Pentagon demands for broader operational flexibility; the outcome will shape how commercial models are integrated into classified and operational systems. If Anthropic accepts revised terms with enforceable safeguards, it could set a hybrid model of departmental authority plus company-implemented safety measures. If the department presses for looser constraints without strong contract safeguards, companies may face intensified reputational and legal scrutiny that could alter their willingness to participate in future defense projects.
Readers should watch for any official joint statement or published contract language after the meeting, which will clarify whether the parties settled on a standard template or whether company-specific exceptions persist. The episode underscores a larger policy choice about where responsibility for preventing misuse should sit: with vendors through technical ‘‘safety stacks,’’ or with purchasers through contractual and operational controls.
Sources
- The New York Times — Reporting on Pentagon-AI negotiations (news media)