The series that translates national-security and AI-policy arguments across partisan lines, because the stakes are too high for tribal shorthand. Quotes below are verbatim and sourced; where a source is reported rather than primary, it is labeled.
What the Emails Actually Showed
The court did the work Congress could not. On July 2, filings in the Northern District of California made the Anthropic-Pentagon negotiation public, and the record does not look like the story either side told in February. The Wall Street Journal reported the documents first; Gizmodo posted the underlying exhibits the same day. What they show is a months-long back-and-forth between Anthropic CEO Dario Amodei and Emil Michael, undersecretary of defense for research and engineering, over one narrow question: what would Anthropic's acceptable-use policy say in a signed contract.
Amodei's redlines were consistent and narrow. Two use cases were off the table: fully autonomous weapons and mass domestic surveillance. Everything else was negotiable, and Anthropic had already agreed to substantial adaptation for defense customers. Michael's counter was that any exclusions were unacceptable. In one January exchange, after Amodei repeated the redlines, Michael answered that the guardrails were, in his own words, "just not workable."
There is no distinction in our world between weapons that are defensive or offensive.
The next day, Hegseth designated Anthropic a supply-chain risk to national security. OpenAI signed a Pentagon deal within hours; CEO Sam Altman later conceded to Axios that the timing "looked opportunistic and sloppy." The March hearing before Judge Rita Lin produced a preliminary injunction blocking the designation; her opinion called the Pentagon's action "classic illegal First Amendment retaliation" for Anthropic bringing the dispute into public view. That injunction is still in effect, and the D.C. Circuit is sitting on a Pentagon appeal after May 20 oral arguments.
What changed this week is that the record is no longer just Anthropic's summary. It is the emails.
What the Procurement Hawks Are Saying
The pro-Pentagon case is not incoherent, and it is worth stating in its strongest form. Defense procurement runs on interoperability, and interoperability runs on standardized terms. When a vendor writes a use restriction into a contract, someone at the acquisition office has to enforce it, in real time, against operational commanders. Michael's line about defensive versus offensive weapons is dismissive on its face, and it is also a description of how deployed systems actually work: a targeting model does not know which way the missile is pointed. From the Pentagon's side, the ask was that Anthropic's models be usable across the mission set the department already runs.
We will not let ANY company dictate the terms regarding how we make operational decisions.
That framing produces a straightforward institutional argument: if a lab wants to be a defense supplier, it has to accept the customer's mission definition or step aside. And the Pentagon has an operational fallback. OpenAI, xAI, Google, and Anthropic itself had all been under $200 million agentic-AI awards since mid-2025; the market did not collapse, and the acquisition pipeline kept moving. To a hawk, that is the point. No single vendor gets to hold the mission hostage.
The honest counter, even from inside the department, is that the "lawful uses" formulation is doing a lot of work. It presumes that every use a lawyer will sign off on is a use the country should want its most capable AI applied to, and that presumption is exactly what Anthropic refused to make in writing.
What the Guardrails Camp Is Saying
Amodei has been consistent from the beginning, and his position in the unsealed emails matches the public letter Anthropic put out February 26. Two uses, and only two, were unacceptable: fully autonomous lethal systems with no human in the loop, and mass surveillance of Americans. Everything else, including partially autonomous targeting of the sort deployed in Ukraine, was on the table.
Regardless, these threats do not change our position: we cannot in good conscience accede to their request.
The Trump administration’s decision to label Anthropic a supply chain risk to national security weakens free enterprise, harms innovation, and makes Americans less safe.
I’m deeply disturbed by reports that the Department of Defense is working to bully a leading U.S. company, which has already provided enormous utility to the intelligence community and warfighter.
Elizabeth Warren opened a formal Senate Armed Services investigation into the designation and into the OpenAI contract that landed the same afternoon. Her staff has been circulating a separate line of inquiry about Emil Michael's financial disclosures. Per Office of Government Ethics filings surfaced by The Lever and ProPublica, Michael sold xAI stock originally disclosed at $500,000 to $1 million on January 9 for between $5 million and $25 million, a return between 400% and 4,800%. The Pentagon later moved to bring xAI's Grok model onto classified networks, work that had previously been earmarked for Anthropic's Claude. Warren's letter to Hegseth pushed him on "what appears to be retaliation by the Department of Defense against artificial intelligence contractors that seek contractual guardrails to prevent the misuse of their AI tools."
None of that is dispositive on its own. Taken together it describes a procurement decision that arrived pre-loaded with a financial motive for at least one senior decision-maker, and the unsealed emails give the retaliation theory a paper trail.
The Provision Nobody Will Own on the Floor
The Senate Armed Services Committee reported S. 4784 out on an 18 to 9 vote, and Bloomberg Government's July 8 story flagged the provision that matters here: new oversight over how the Secretary of Defense designates suppliers as national-security risks. The details of the closed markup are not public, but the direction is. The committee added process where none existed, and it did so on a bipartisan majority. Bloomberg's Roxana Tiron characterized the provision as the first legislative step lawmakers have taken in reaction to the Pentagon-Anthropic fight, and noted it signals the committee is not fully behind the department's use of the designation power. Underlying report language will publish with the file copy.
The other half of the response is investigative. Warren's inquiry, Coons's public statement, and Warner's floor posture point in the same direction: the designation itself, not any Anthropic policy, is the concern. And this is where the alignment gets uncomfortable for anyone selling a partisan version of the fight. Republicans on Armed Services voted to advance a bill that includes the oversight provision. National-security Democrats voted for the underlying authorization. The Pentagon does not have unanimous backing for how it handled this, and the record now shows why.
On the other side of the aisle, the procurement-hawk position has a real constituency and a real point. Congress does not run acquisition, and neither does the general counsel of any single lab. If Anthropic gets to write its own use restrictions into every contract, so does every other supplier, and the department loses the ability to standardize. That is a legitimate concern, and it is exactly what the Senate provision tries to solve by requiring a documented process for the designation rather than a tweet from the secretary's account.
Anna's Read: The Redline That Became a Contract Term
For years, "lawful uses" was procurement shorthand for a whole set of assumptions nobody wrote down. Fully autonomous lethal decision-making was assumed to be off the table because of Directive 3000.09 and its human-judgment requirement. Domestic mass surveillance was assumed to be off the table because of the Fourth Amendment and executive orders that predate this administration. Anthropic's contribution was to notice that neither assumption is still safe, and to insist that the contract say what the department's own guidance used to guarantee.
That is a defensive move for the audience I write for. For CISOs at frontier labs, it means the acceptable-use policy is now a procurement instrument, and every future government contract will negotiate against it in the open. For DIB program offices, it means a supply-chain-risk designation is a live threat you have to plan around, and the Senate provision is the first hint you may get formal notice before an OSD press release lands. For congressional staff on Armed Services and Intel, it means the FY 2027 NDAA conference is the venue where this fight actually gets decided, and floor amendments to strip the oversight language will be the tell for who thinks the Pentagon's original move was correct.
My read is straightforward. The redlines Anthropic held are the ones a competent procurement lawyer should have written in on the government's side. The Pentagon's insistence that they were "not workable" is a description of how far the operational envelope has drifted from the policy envelope, not a defense of the drift. And the Senate's response, small as it is, is the right shape: put process around the designation power, so the next time it gets used it is used on a supplier and not on a policy dispute the department lost.
What to watch: whether the S. 4784 oversight provision survives conference with the House NDAA; whether the D.C. Circuit rules on the Pentagon appeal before the recess; whether Warren's Emil Michael inquiry produces sworn testimony; and whether any Republican senator publicly defends the designation on the merits rather than the procurement authority. That last one is the tell. If nobody will say the designation was correct, the precedent does not survive the fiscal year.
Related Briefings
Anna R. Dudley writes on national security, intelligence policy, and the places where hawks and reformers need to find each other. Bipartisan Translation is the weekly series for the conversation that is not happening on cable news. Subscribe at annardudley.substack.com.