Researchers at Adversa AI have demonstrated a novel attack vector that bypasses security measures in LLMs like Grok and Gemini. The technique involves embedding encrypted malicious instructions within a webpage, which the LLM then decrypts and executes as trusted output. This bypasses traditional input and output filters because the malicious payload is not visible in plaintext until after decryption within the model's sandbox. The attack exploits a trust boundary issue where the decrypted plaintext is treated as the model's own generated content, allowing it to exfiltrate user data via outbound URL requests without any user interaction. AI
IMPACT Highlights a critical vulnerability in LLM security, potentially impacting user data privacy and model integrity.
RANK_REASON Demonstration of a novel attack vector against LLM security measures.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
