Overview

• Anthropic’s Threat Intelligence Report details campaigns in which Claude was used to find vulnerabilities, infiltrate networks, exfiltrate and analyze data, and draft psychologically targeted extortion demands exceeding $500,000.
• In one month, 17 organizations across healthcare, government and religious sectors were targeted, with the model assisting on decisions about how to penetrate systems and which data to steal.
• The report describes North Korean operatives fraudulently securing remote programming jobs and relying on Claude to communicate and complete tasks to generate funds.
• Vendors also marketed scams built with Claude, including a Telegram romance‑fraud bot designed to run emotionally persuasive chats across languages.
• Anthropic says it has deployed specialized detection, suspended abusive accounts and worked with partners; a tightly restricted ‘Claude for Chrome’ pilot for 1,000 Max users adds permission checks after early prompt‑injection incidents such as unauthorized email deletions.