Anthropic Says Russian Chinese Actors Used Claude for Weapons Development
Left 50%
Center 0%
Right 50%
1 left · 0 center · 1 right
What happened
On Thursday, U.S. AI company Anthropic published a report saying it had identified and banned accounts that used its Claude model for weapons-related work by actors in Russia, China, northern Yemen, and an unspecified Iran-nexus group, and that it had shared information with government authorities and strengthened safeguards. A small Russian group tied to a regional university—which Anthropic said was not a Russian state entity—used nine accounts created between late 2025 and early 2026, routing around geographic controls via commercial virtual private servers, to build drone-swarm software including shared swarm memory, fault-tolerant coordination, and systems Anthropic said could select targets and issue detonation commands without a human in the loop; the group tested the software on hardware, trained a tool on scraped Ukrainian combat footage, and used a fixed coordinate in Donetsk as a demonstration strike point. A China-based actor used Claude to write Chinese-language technical specifications and a more than 200-page anti-torpedo fire-control proposal aimed at the Chinese navy, comparing the design with publicly described U.S. programs, while another China-based user gathered intelligence on directed-energy weapons. Actors in northern Yemen used Claude Code for guidance, navigation, and control software for rockets and missiles, including work Anthropic described as a guided rocket with a commodity phone-class flight computer, a multi-stage ballistic missile, and a multi-variant missile with a hypersonic-glide-vehicle variant; Anthropic said it had no evidence they fielded an operational device but that they test-fired a guided rocket that failed. An Iran-nexus actor used Claude to compile publicly available targeting material on U.S. naval forces, including personnel names scraped from photo captions, ship and aircraft transponder identifiers, commercial satellite-imagery query scripts, and websites that exposed U.S. naval movements.
Sourcing check · Fully on the record · 2 articles
2 of 2 articles cite a named source for the story. It was on the record from the first report.
What these terms mean and how they're counted
- On the record · names a source 2 of 2
- On the record means the article cites a source by name for the story's main claim: a person, or an institution speaking officially, such as a company statement or a named agency's spokesperson. The opposite is an unnamed source. The count is how many of the counted articles cite at least one such source; the names are everyone named across them, each once however the articles spell it. Here: Anthropic
- Primary evidence · cites a document or data 2 of 2
- Primary evidence means the article cites a record you could check yourself: a court filing or ruling, a dataset or report, a recording or transcript, an official statement or press release, market data. A quote is not a document. The count is how many articles cite one. Here: company report
- Credits another outlet 0 of 2
- The article says the story comes from another outlet's reporting (“first reported by Reuters”). Those articles relay the story; they are not independent confirmation. Here: Each article presents its own reporting.
- Unnamed sources 0 of 2
- The main claim is attributed to people who are not identified: “people familiar with the matter”, “officials said”. The count is how many articles do this.
- How it developed
- When the first report landed and whether it named anyone; when the first named source and the first document appeared. Here: First report Sep 11, 15:30 UTC by Business Insider, naming its source. First document in the first report, from Business Insider (company report).
- Not counted 0
- Opinion, analysis and video pieces, and articles too short to assess. Every count above is out of the counted articles. An article republished under a new headline is one article.
This is about who is cited, not whether the story is true. Each source below says what it names. How we count
Omitted — what each side leaves out
Unpacked
Right-leaning coverage never tells readers that Russian actors used Claude to write drone software that could pick targets and issue detonation commands on its own, then ran that software on actual hardware. Left-leaning coverage in Business Insider describes a specialized freelance group tied to a regional university that built swarm coordination and designed the platform for "autonomous lethal engagement." The New York Post folds Russia and China into a passing note about individuals drafting plans for conventional weapons.
Left-leaning coverage opens on Russian and Chinese threat actors and drone swarming. Right-leaning coverage opens on Iran-backed rebels in Yemen trying to build missiles, and only later mentions Russia and China.
Did any Claude-assisted weapon ever fly in combat, or only in a failed rocket test and a hardware trial?
Bottom line
A New York Post reader would take Anthropic's report as a failed Houthi missile experiment. That outlet never reports the Russian hardware-tested swarm that Business Insider says could detonate without a human in the loop.
The receipts — 5 facts one side's coverage lacks
- Russian actors used Claude to build drone-swarm software that could select targets and issue detonation commands without a human in the loop, then tested it on actual hardware
- Anthropic identified nine accounts that circumvented geographic access controls by routing traffic through commercial virtual private servers
- An Iran-nexus threat actor used Claude to build an identity-profiling harness targeting Israeli governmental and non-governmental individuals and Jewish diaspora organizations
- Yemen-based actors worked on a multi-stage ballistic missile and a hypersonic glide vehicle variant
- A China-based actor used Claude to compile a more than 200-page proposal for an anti-torpedo fire-control system
The Left View
Business Insider framed the report as providing examples of a frontier American AI model being used to advance adversarial military capabilities, situating it as the Trump administration and Silicon Valley navigate AI development amid security and safety concerns. The coverage stressed Anthropic’s finding that the Russian actors designed a platform for “autonomous lethal engagement” in which “the onboard model could select targets” and “issue detonation commands without a human in the loop,” and noted that Anthropic identified the activity through its own investigations, banned the accounts, and strengthened safeguards. It also linked the report to researcher Jacob Coxon’s resignation days earlier in protest of the industry’s development race, quoting his warning that “the people building AI earnestly believe that it could kill us all by the end of the decade.”
The Right View
The New York Post framed the report as “alarming” and led with Iran-backed Houthi rebels trying to use Claude to craft software for ballistic missiles, stating that Anthropic’s unnamed “cell of threat actors based in northern Yemen” almost certainly refers to the Houthis, who control the region. It highlighted that the actors hid their goals and split requests across sessions, and that safeguards “blocked many of their requests, but not all of them,” and it also flagged an Iran-nexus “identity-profiling harness targeting Israeli governmental and non-governmental individuals, and Jewish diaspora organizations,” disrupted biological-weapons research, and Russia- and China-linked propaganda and espionage cases. The coverage said the warnings will fuel fears about weaponized AI and cited Coxon’s claim that Anthropic and OpenAI are “racing straight to self-improving superintelligence and gambling with our lives.”
Our Take (balanced)
The left-leaning account’s strongest argument is that a U.S. frontier model was used to design autonomous lethal drone systems that could select targets and detonate without a human in the loop—Anthropic’s own description of that platform, hardware testing, and a Donetsk demonstration point is the evidence—and that this sits inside an industry race whose own researchers warn could be existentially dangerous. The right-leaning account’s strongest argument is that Iran-backed Yemen-based actors and other hostile-state-linked groups are already using commercial AI as a stand-in for weapons engineers and intelligence analysts against U.S. and Israeli targets, with the failed guided-rocket test-fire, the missile-variant work, the U.S. naval targeting compilation, and the Jewish-diaspora profiling harness as the evidence. The central unresolved tension is whether the story is primarily about frontier-model capability enabling autonomous weapons and runaway industry risk, or about adversarial states and proxies exploiting Western AI products for conventional military and intelligence ends.
2 sources
More stories
The week's bottom lines, in your inbox
One email a week: the five stories that mattered and what they actually mean. Free.