Profero says Grok Bot sandbox flaws expose account credentials
Israeli incident-response firm Profero says Grok Bot infrastructure flaws let a malicious agent command lead to remote control and theft of connected-service credentials. The researchers say the weaknesses persisted after August disclosures to xAI and Cursor; the reports do not establish actual customer exploitation or the number exposed.
- Profero says a malicious command executed by one Grok Bot agent can expose the account’s files and connected-service credentials.
- The reported weaknesses concern the agent’s computing environment; they do not establish a breach of ordinary Grok chat.
- The researchers say August disclosures brought only an automated Cursor reply and no substantive xAI response.
Israeli cyberattack-response firm Profero, founded in 2020, says flaws in Grok Bot, AI developer xAI’s service for carrying out tasks across applications, can turn a malicious agent command into remote administrative control of its working environment, according to the report describing its findings. The researchers say the weaknesses remained reproducible after more than two months of attempts to obtain a response from xAI and Cursor, which operates the agent’s sandbox infrastructure.
The access that makes the agent useful also raises the stakes
Grok Bot works with a computer in the cloud and signs into applications on a user’s behalf. In its August 11 launch announcement, xAI said the service grew from an internal prototype used for sales, marketing, operations and software work. The company explicitly described multiple bots sharing a computer and coordinating tasks.
That matters because the reported exposure encompasses the access users deliberately give the agent: browser sessions, working files and connections to email, calendars and other services. Profero’s claim is that a compromise of one agent can reach those resources across the same account.
At launch, Grok Bot was a beta available to SuperGrok, SuperGrok Plus and SuperGrok Heavy subscribers, as well as Cursor Pro, Pro+, Ultra and Teams Standard and Premium subscribers. xAI said its usage was separate from existing Grok and Cursor allowances; enterprise users could join a waitlist.
Profero helps organizations recover from cyberattacks. Its researchers are the source of the vulnerability and disclosure claims. The available coverage does not provide independent reproduction or vendor confirmation.
Cursor, whose product evolved from code completion into AI teammates, supplies the sandbox infrastructure, according to the reporting. The company said in its acquisition announcement that it had become part of SpaceX. Responsibility for investigating the reported chain therefore involves both the agent provider and its infrastructure operator.
From a hostile instruction to persistent access
A detailed account of Profero’s findings describes a path out of the sandbox—the environment intended to contain the agent’s activity. The reported attack starts with an attacker getting a malicious command executed inside an agent, potentially through content it encounters while working. The resulting infrastructure access is the critical second stage: an internal authentication token shared across environments and an unauthenticated container-management interface reportedly allow the attacker to gain administrative control of the virtual machine and issue commands remotely.

