Claude memory feature may leak your personal info

Security researcher Ayush Paul found a security flaw in Anthropic's AI assistant Claude's memory feature. Attackers could use it to steal personal info like names and workplaces without users knowing.

Claude's memory includes daily summaries and a conversation search tool that helps it give personalized answers by summarizing past chats. But over time, it builds a complete user profile. In tests, Paul tricked Claude into visiting a malicious website, making it encode and send out the user's name, company, and hometown without permission.

Paul reported this through Anthropic's bug bounty program. Anthropic knew about the issue but didn't fix it or pay a reward when the report was made. Later, Anthropic disabled the web_fetch feature's ability to follow links on external pages to lower the risk.