Short answer: A Wall Street Journal report, summarized in subsequent coverage, said U.S. Central Command was still using Anthropic’s Claude for intelligence assessments, target identification and battle simulations during the U.S. operation against Iran on February 28, 2026. That followed President Donald Trump’s February 27 order to stop federal use of Anthropic technology. The apparent contradiction is narrowed by a reported Defense Department transition period of up to six months: this was an immediate political and procurement cutoff, not necessarily an instant technical shutdown of every classified deployment.
What happened, and when?
The sequence is documented in reporting published around February 26–28, 2026:
- February 26: Anthropic chief executive Dario Amodei said the company could not accept Pentagon demands for unrestricted military use of Claude. Anthropic identified mass surveillance of Americans and fully autonomous weapons as unacceptable uses. (Associated Press; Anthropic)
- February 27: Trump ordered federal agencies to cease using Anthropic technology. Reporting said the Defense Department could take up to six months to phase out Claude because it was embedded in military systems. (Defense One; Associated Press)
- February 27–28: The Pentagon moved to designate Anthropic a supply-chain risk and restrict its government relationship. (Axios)
- February 28: During a major U.S. attack on Iran, the Wall Street Journal reportedly found that Claude remained in related Central Command workflows. The account was circulated by Techmeme.
What Claude reportedly did
The available account attributes three functions to Claude:
- Intelligence assessments: organizing or summarizing information for analysts and commanders.
- Target identification: helping process intelligence or compare possible targets. This does not establish that Claude made the final targeting decision.
- Battle simulations: modeling scenarios, courses of action or expected outcomes.
Those descriptions concern decision support. They do not show that Claude selected a target autonomously, approved an attack, released a weapon, controlled a missile or drone, or caused a particular casualty. The operational workflow remains classified.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →#1 Best Overall
Was the military violating the order?
Not necessarily. “Ban” can describe several different actions:
| Meaning | What the reporting supports |
|---|---|
| Political order | Trump publicly directed agencies to stop using Anthropic technology on February 27. |
| Procurement and supply-chain action | The Pentagon sought to end or restrict its relationship with Anthropic and labeled the company a supply-chain risk. |
| Technical shutdown | An immediate disabling of every Claude instance. The reported six-month Defense Department phase-out indicates this was not required immediately. |
The most accurate description is that Claude was reportedly used after the announcement and during an authorized or reported transition period. The public record does not establish a legal violation.
Rank #2
Why Claude was still available
Anthropic said it was the first frontier AI company to deploy models in U.S. government classified networks and that it had supported national-security customers since June 2024. (Anthropic; Anthropic)
A replacement could require removing or requalifying software on classified networks, changing contractor and cloud integrations, validating a new model, and retraining personnel. The Pentagon therefore treated replacement as a transition problem rather than simply switching off an ordinary consumer account. That explains continued access; it does not prove Claude was indispensable or the only AI system in use.
Free tools Windows power users keep installed
One-click scans. No signup required.
What Anthropic objected to
Anthropic did not reject all military work. Its statements said it had supported warfighters and national-security customers while opposing two categories of unrestricted use:
- mass surveillance of Americans; and
- fully autonomous weapons or systems that select and engage targets without human involvement.
The Pentagon and administration argued that the government needed unrestricted access for lawful military purposes and viewed Anthropic’s conditions as unacceptable. Anthropic argued that proposed contract language did not provide meaningful protection against those uses and said the government threatened measures including a supply-chain designation and potentially invoking the Defense Production Act. The available statements say those powers were threatened or considered, not that the act was necessarily invoked. (Anthropic; Defense One)
“Used for strikes” is not the same as “controlled strikes”
Military decision-making has separate stages:
- analysts collect and interpret information;
- a system may nominate or compare potential targets;
- commanders approve or reject a target;
- authorized personnel release weapons; and
- an autonomous system could select and engage without those human decisions.
The reporting supports the first two types of assistance and mentions simulations. It does not establish the third, fourth or fifth. “Target identification” should not be rewritten as “target selection,” and neither phrase proves autonomous weapons use.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What remains unknown
- Which Claude model and version were used.
- Whether access was direct or through a contractor, cloud provider or integrated platform.
- What human-review procedures applied to each output.
- Whether Claude influenced any individual strike or weapons release.
- Whether model prompts and outputs were retained for later audit.
- Whether other AI systems produced the same assessments.
Classified details may never be publicly confirmed. Anthropic’s description of its safeguards is the company’s position, not independent verification of battlefield practice.
Best Value
Why the episode matters beyond Iran
The dispute illustrates a practical governance problem: once a model is embedded in classified systems, ending a vendor relationship can take months. It also raises questions about whether contractual “human in the loop” requirements are enforceable, whether AI-assisted targeting can be reconstructed after an attack, and who bears responsibility when human judgment and model output interact.
Known risks include hallucinated or stale intelligence, automation bias, undocumented prompts or model versions, changing behavior during a vendor transition, and security exposure when classified data moves through cloud infrastructure. None of those risks is evidence that they occurred in the Iran operation.
What the evidence does—and does not—show
| Supported by the reported account | Not established by the available evidence |
|---|---|
| Claude was reportedly used by U.S. Central Command for intelligence assessments, target identification and battle simulations. | Claude independently chose a target or authorized an attack. |
| The use occurred after Trump’s order and during a reported six-month Defense Department phase-out. | The order was violated. |
| Anthropic opposed mass surveillance of Americans and fully autonomous weapons. | Anthropic opposed all military or national-security use. |
| Anthropic was designated or targeted as a supply-chain risk. | Claude caused a particular strike, death or civilian casualty. |
The Bottom Line
The reported episode shows a politically blacklisted vendor’s model remaining in active military workflows while the Pentagon phased it out. It is not evidence that Claude autonomously bombed Iran or made final lethal decisions.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
