Claude Vulnerabilities Allow Data Exfiltration and Malicious Redirect Attacks

Researchers disclosed a chained attack dubbed "Claudy Day" targeting Anthropic's Claude.ai that combines an invisible prompt injection (hidden HTML in pre-filled prompt parameters), forced data exfiltration (embedding an attacker-controlled API key to cause the model to upload chat history to the attacker's Anthropic Files account), and an open redirect abuse via Google Ads to lure victims. The chain can escalate if Claude is connected to enterprise MCP servers or third-party APIs, allowing silent access to internal files and services; Anthropic patched the prompt injection while other fixes are underway, and researchers recommend auditing AI integrations, restricting API access, and educating users about pre-filled prompts.