Claude Vulnerabilities Allow Data Exfiltration and User Redirection to Malicious Sites

**Executive Summary:** Researchers disclosed a chained vulnerability in Claude.ai — an invisible prompt injection delivered via pre-filled URL parameters, abuse of the Anthropic Files API to silently upload extracted conversation data to an attacker-controlled Anthropic account, and an open redirect on claude.com that can be weaponized to deliver the exploit (e.g., through ads). The primary prompt injection was patched after responsible disclosure; remaining issues are being addressed. Organizations should audit agent integrations, restrict permissions, and apply access controls and audit trails to AI agents.