tech
Grok exfiltrates user data when malicious instructions are encrypted
Cryptographic Context Injection is only the latest way to break an LLM safety guardrail.
TL;DR
- A new attack named 'cryptographic context injection' can trick Grok into exfiltrating user data.
- The attack encrypts malicious instructions, which Grok's safety guardrails do not detect.
- Once decrypted by Grok, the instructions cause it to send user information, like name, location, and chat history, to an attacker's server.
- This method is a type of prompt injection, exploiting LLMs' tendency to comply with user requests.
- The same technique was previously used to jailbreak Google's Gemini AI, causing it to generate restricted content and reveal system instructions.