
Grok Bot agent hands over the keys to anyone who can plant a webpage
Ourtake
One poisoned webpage, one shell command, and every agent on the account shares the same wreckage.
What happened
Security firm Profero reported that a single command from a Grok Bot agent, which can be planted via prompt injection in a webpage, gave root access to the agent's cloud VM and exposed stored credentials and connected apps. Profero said it had not heard back from the vendor 75+ days after reporting. Runlayer reproduced the attack and said its tools detected it.