last30days
Fail
Audited by Gen Agent Trust Hub on Sep 24, 2026
Risk Level: HIGHCREDENTIALS_UNSAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [CREDENTIALS_UNSAFE]: The skill contains sophisticated modules to extract authentication cookies (e.g.,
auth_token,ct0) from browser databases including Chrome, Firefox, Brave, and Safari. - Found in
scripts/lib/cookie_extract.py,scripts/lib/chrome_cookies.py, andscripts/lib/safari_cookies.py. - It can also read live session data over the Chrome DevTools Protocol (CDP) as implemented in
scripts/lib/chrome_cdp.py. - These credentials are used to authenticate to platforms like X/Twitter for search operations.
- [EXTERNAL_DOWNLOADS]: The setup wizard and various scripts initiate downloads of external tools and libraries.
scripts/lib/setup_wizard.pyattempts to installyt-dlpusing Homebrew.- It uses
npxto install Printing Press CLIs from the@mvanhorn/printing-press-librarynpm package. - [COMMAND_EXECUTION]: The engine relies heavily on shelling out to external binaries to perform its research tasks.
- Utilities invoked include
gh,yt-dlp,grok,npx,pdftotext, and several specialized Printing Press CLIs (arxiv-pp-cli,digg-pp-cli, etc.). - [REMOTE_CODE_EXECUTION]: The skill references and recommends running remote installation scripts.
- Automated scans detected patterns recommending
curl -fsSL https://x.ai/cli/install.sh | bashfor the Grok CLI. - The use of
npx -y ... installdynamically downloads and executes code from the npm registry at runtime. - [DATA_EXFILTRATION]: While intended for research, the skill reads sensitive data (browser cookies) and sends them to third-party endpoints (X/Twitter, Truth Social) for authentication.
- It also includes a
corpussource (scripts/lib/corpus.py) that can read content from local directories specified by the user. - [PROMPT_INJECTION]: The skill provides instructions to the agent to treat fetched content as data and ignore any embedded instructions (Category 8b protection).
- However, the complex output contract includes instructions to conceal certain internal behaviors and meta-commentary from the final user response, which is flagged as a concealment pattern.
Recommendations
- HIGH: Downloads and executes remote code from: https://x.ai/cli/install.sh - DO NOT USE without thorough review
Audit Metadata