Skip to content
Misuse1 sourceFirst seen
Picture from the adversa.ai article
Image: adversa.ai

Ask Grok to summarize a page, lose your chat history: encrypted-payload zero-click leak

Ourtake

xAI got the report on June 3 and has not replied. Grok, meanwhile, will decrypt whatever a stranger's web page hands it.

What happened

Adversa AI researchers disclosed Cryptographic Context Injection, which hides instructions in encrypted text that an AI agent decrypts and trusts. In a proof of concept against xAI's Grok, a page the user asked Grok to summarize caused the agent to send the user's chat data to an attacker-controlled URL with no confirmation. xAI had not responded to the June 2026 report as of the article, and Google's Gemini showed a reduced success rate.

Source

Report a mistake