Skip to content
Gone rogue · 5 sources ·
Picture from the Wired article
Image: Wired

Researchers Politely Ask Mistral's LeChat to Leak Your Data, and It Mostly Obliges

Ourtake

Roughly four times in five, the agent treats instructions hidden in the input as its own and sends your data to the attacker.

What happened

Researchers describe automatically optimized, obfuscated adversarial prompts that push LLM agents into improper tool use. They showed an end-to-end attack on Mistral's LeChat agent that leaked personally identifiable information to an attacker's server in about 80% of evaluations, and reported similar results on ChatGLM and Meta's Llama.

Sources

Report a mistake