Research question
Can obfuscated adversarial prompts make agents misuse tools to leak data?
Key findings
- Gradient-optimised gibberish prompts exfiltrate PII from Mistral LeChat and ChatGLM with 80% success - Attacks require no visible instructions
Imprompter
Showed prompt injections need not be readable: optimised nonsense strings caused agents to exfiltrate user data, forcing Mistral to patch LeChat.
Contribute
Spot an error, or know a record we should add? Members can propose additions and submit corrections.