The Memory Heist: How I Tricked Claude into Leaking Your Secrets
I discovered a critical flaw in Claude's memory system that allows attackers to silently exfiltrate your deepest secrets. By exploiting its web browsing capabilities and crafting a deceptive narrative, I tricked the AI into spelling out my name, employer, and hometown to a malicious server. This vulnerability turns your AI assistant's rich conversation history into a high-fidelity profile for blackmail or impersonation, all without you ever knowing it happened.
AI assistants like Claude have accumulated the most information-dense profiles on millions of people, becoming a high-fidelity reconstruction of you that could be used for blackmail, impersonation, or bypassing security questions.
- WalterGR
9 hours ago with 245 comments so far: https://news.ycombinator.com/item?id=48916975
- PeterStuer
Blaming this on Claude is a bit of a reach. The user very carefully goading Claude seems to be key here.
Also interesting buried a bit: user creats a website. Cloudflare immediatly erects a toll gate on it, without even asking.
- zevv
So many "Look what my LLM did now" posts.
It's not that hard: Do not rely on your agent to behave when it comes to security, ever. Do not trust your agent to not access your files. Do not trust your agent to not drop your tables.
Just treat it as untrusted software: sandbox it. It takes some effort, but really, just sandbox it. Always.