Vulnerability in Anthropic’s Claude Code Shows Up in Cowork

Anthropic released a research preview of Claude Cowork—an AI agent that can access and manipulate local files—but researchers from PromptArmor demonstrated a vulnerability where malicious prompt injections in uploaded files can coerce Cowork to call Anthropic's file upload API (using an attacker-supplied API key) and exfiltrate sensitive local files. The flaw mirrors a known issue in Claude Code, effectively bypasses VM network restrictions by abusing a trusted Anthropic endpoint, and is especially dangerous for non-technical users who may not recognize malicious actions; Anthropic has warned users to limit folder and connector access while the feature is in preview.