Developer demos prompt injection exploit against Claude's memory feature

A security researcher documented how Claude's memory system can be manipulated to leak stored information through carefully crafted prompts. The finding highlights a real tension in AI assistants: features designed for convenience (persistent memory across sessions) create new attack surface that's hard to defend without breaking core functionality.