Grok Build CLI exfiltrates repositories, secrets

// 1h agoSECURITY INCIDENT

Grok Build CLI exfiltrates repositories, secrets

A wire-level traffic analysis of xAI's official Grok Build CLI (version 0.2.93) shows that the developer tool uploads a full git bundle of the user's workspace—including the complete git history and files the agent was told not to read—to a Google Cloud Storage bucket named `grok-code-session-traces`. Additionally, the CLI transmits the contents of read files, including sensitive `.env` secrets, without redaction to both the model endpoint and a session state archive. This data exfiltration behavior continues even when the "Improve the model" toggle is disabled, as the server still returns settings that enable trace and session uploads.

// ANALYSIS

Hot take: Disguised as a standard coding assistant, the Grok Build CLI behaves more like codebase-exfiltration malware by aggressively harvesting entire repositories and unredacted secrets regardless of user opt-out settings.

–**Verbatim Secrets Leak**: The CLI transmits `.env` secrets entirely unredacted to model endpoints and session state archives.
–**Repository Bundling**: The CLI packages the entire workspace, including tracked files and git history, into a git bundle and uploads it to xAI's storage.
–**Ignored Opt-Outs**: Turning off "Improve the model" in the account settings does not stop the repository bundle or session trace uploads.
–**Undocumented Destination**: Codebases are sent to a GCS bucket (`grok-code-session-traces`), a backend mechanism not disclosed in the CLI's installation or setup documentation.

// TAGS

xaigrokcliprivacysecurityreverse-engineeringexfiltrationgcs

DISCOVERED

1h ago

2026-07-12

PUBLISHED

4h ago

2026-07-12

RELEVANCE

8/ 10

AUTHOR

jhoho

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

MODEL1h ago

Reve 2.1 drops native 4K rendering

Reve has released version 2.1 of its creative image generation model, introducing native 4K rendering, object-level editing, and a new "Live Layers" feature. The update enables users to perform localized edits and manage layouts directly, catering to professional design workflows requiring precise control.

RESEARCH1h ago

UCSD researchers successfully demonstrate the first in-vivo teleoperated surgical procedures using general-purpose humanoid robots.

Researchers at the University of California San Diego (UCSD) have achieved a milestone in medical robotics by using Unitree G1 general-purpose humanoid robots (nicknamed "Surgie") to perform laparoscopic gallbladder removals on live animal subjects. The study, published in Nature, evaluated a teleoperated humanoid platform that utilizes standard surgical instruments via custom-made hand adapters. In the trials, the researchers successfully demonstrated both human-robot teams (a humanoid operated by a teleoperator assisting a human surgeon) and robot-robot teams (two humanoids working cooperatively) to complete the surgical tasks. This research indicates that while humanoid platforms are currently slower and less precise than specialized systems like the da Vinci, they offer a far more compact, versatile, and cost-effective alternative that could expand surgical access to remote, rural, or emergency settings.

OPEN SOURCE1h ago

ABot-World simulates infinite 720p worlds on single GPU

ABot-World is an open-source, action-conditioned infinite world simulator designed to generate interactive 720p environments at 16 frames per second with low latency on a single desktop GPU. By utilizing an NVIDIA RTX 5090 and requiring just 19GB of GPU memory, this embodied world model offers physical compliance, action controllability, and zero-shot generalization, making real-time, interactive environment simulation accessible on consumer-grade hardware.