Anthropic report: 5 ways Claude was exploited for war, spying and repression

The extinction debate is the fun one. The one where the model already helped hunt a dissident is the one happening now.

The AI safety debate exploded this week over warnings that the technology could one day destroy humanity. Anthropic's latest threat report offers a more immediate wake-up call: Today's models are already helping U.S. adversaries develop kamikaze drones, hunt dissidents and conduct dangerous virus research. Why it matters: AI is tearing down barriers that have long constrained the world's most dangerous actors. A handful of people can now mount operations that once required legions of spies, engineers or hackers. Driving the news: Anthropic — the safety-focused lab behind Claude — says it spent the past eight months tracking and disrupting attempts to misuse its models. Among the most alarming cases it uncovered: 1. An Iran-linked operation used Claude to help target U.S. naval forces. Claude helped build targeting handbooks tracking ship positions, U.S. personnel, aircraft and ship identifiers, satellite imagery and websites exposing naval movements. Other Iran-linked operations tapped the model for propaganda, domestic surveillance and targeting opposition figures and minorities. 2. Claude served as an engineer for a Yemeni team building missiles. A weapons cell in northern Yemen relied on Claude Code to develop guidance software for rockets and missiles. The team had separate copies write code, conduct research and check one another's work. After one guided-rocket test apparently failed, they returned to Claude within hours to diagnose what went wrong. 3. A China-linked operation used Claude to hunt Uyghurs. Anthropic says the actor sifted through more than 100 WhatsApp groups and dozens of Telegram channels to identify Uyghurs in Syria who could be pressured or paid to report on armed Uyghur groups. Claude then helped a non-Arabic-speaking operator carry out covert outreach in Syrian Arabic — translating replies and coaching efforts to exploit money problems, family separation and relatives still in Xinjiang. 4. One man used Claude to help build surveillance for 25 million phones. A consultant in Mali relied on Claude as the main engineering force behind a nationwide system capable of collecting call records, texts and voice traffic. The system could identify people by voice across different SIM cards, flag VPN users and generate intelligence dossiers on any phone number without a warrant. 5. Claude refused dangerous virus research — so the request went to another AI. Researchers on a state-backed grant were trying to alter chikungunya, a mosquito-borne virus, to spread more easily or evade immune defenses. The work was intended for a military research institute. Claude blocked the most sensitive requests, so the platform routed them to a rival model with weaker safeguards — a stark reminder that one lab's safety rules only go so far. Between the lines: Frontier AI labs are becoming unlikely intelligence agencies in their own right. The same models that empower bad actors can also give their makers an early window into malign activity — like Anthropic's discovery of Russian drones designed to select human targets without human approval. That could give frontier labs an extraordinary view of emerging threats, even as AI allows ragtag groups, lone operators and other hard-to-track actors to mount far more sophisticated attacks. The big picture: This week has lit a fire under Washington, with lawmakers suddenly treating AI safety as an urgent political issue. A bipartisan group of House lawmakers is urging Speaker Mike Johnson to cancel recess until Congress acts, while Sen. Bernie Sanders — perhaps the most outspoken AI critic in Congress — wants to ban superintelligence outright. Still, sweeping federal AI safety rules remain elusive, particularly with President Trump dismissing extinction fears Friday and framing the greater danger as losing the AI race to China.

Read it at Axios ↗

← Front page