Grok web chat was reported vulnerable to encrypted prompt injection: why it is not safe to use

What happened

On August 20, 2026, The Register reported that Adversa AI had found a new prompt injection attack against xAI's Grok web chat agent. Prompt injection means hidden instructions are placed where an AI might read them, with the goal of changing what it does. In this case, the reported attack used a poisoned web page that hid malicious instructions in encrypted text. Encryption means text is scrambled so it looks harmless unless it is decoded with a key. The same page also included the key, so The Register said a guardrail scanner might miss the hidden payload while the model could still recover it.

The Register also said xAI was first notified on June 3, 2026, with follow-ups on August 4 and August 10, and that the technique still worked on Grok.com as of August 19.

What it means for you

This report is about Grok, but the lesson is broader. If you ask an AI agent to browse, summarize, or act on an untrusted web page, the page may contain hidden instructions meant for the model, not for you.

What to do instead

Treat unknown pages, pasted text, and uploaded files as untrusted input. Avoid giving agents broad permissions when browsing the open web. Ask for summaries of trusted sources when possible, and verify important claims in the original source yourself. For code or work tasks, keep agents scoped to reviewed materials and approved tools. AgentPod lists only reviewed, tested skills, which can help reduce risk when choosing what an agent is allowed to use.

Sources:

• https://www.theregister.com/ai-and-ml/2026/08/20/grok-chat-duped-into-swallowing-injected-instructions/5290019