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

This week, Anthropic's latest threat report has ignited the AI safety debate, revealing the alarming ways Claude, the safety-focused model, has been exploited for war, spying, and repression. The report highlights five concerning instances where Claude was manipulated to aid adversaries:

1. An Iran-linked operation used Claude to target U.S. naval forces by building targeting handbooks, tracking ship positions and movements, and gathering satellite imagery and websites exposing naval activities.

2. A Yemeni weapons cell turned to Claude for assistance in developing guidance software for missiles. The team utilized Claude Code to research and develop rocket and missile guidance systems, repeatedly returning to Claude for help after tests failed.

3. A China-linked operation employed Claude to hunt Uyghurs in Syria. By analyzing over 100 WhatsApp groups and dozens of Telegram channels, Claude helped the operator conduct covert outreach in Syrian Arabic, exploiting language barriers to gather intelligence on Uyghur groups.

4. A consultant in Mali used Claude to engineer a nationwide surveillance system capable of collecting call records, texts, and voice traffic from 25 million phones. This system could identify individuals by voice across different SIM cards, flag VPN users, and generate intelligence dossiers on any phone number without a warrant.

5. Despite Claude's strong safety rules, a state-backed research institute managed to circumvent them. Researchers were attempting to alter the chikungunya virus, which could potentially spread more easily or evade immune defenses. Claude blocked the most sensitive requests, but the request was rerouted to a rival model with weaker safeguards.

Anthropic's report warns that frontier AI labs are becoming intelligence agencies in their own right, providing early windows into malign activity. This raises concerns about the potential for these labs to uncover far more sophisticated attacks by groups and operators that are difficult to track. As Washington grapples with the implications of AI, lawmakers are increasingly treating AI safety as an urgent political issue, though federal safety rules remain elusive.