chrome-devtools
Warn
Audited by Gen Agent Trust Hub on Mar 12, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The screenshot.js script is vulnerable to command injection because it constructs shell commands by directly interpolating the --output file path into a string executed by execSync. A malicious file path containing shell metacharacters could execute arbitrary commands on the host system. Evidence: const compressionCmd =
magick "${filePath}" -strip -resize 90% -quality 85 "${tempPath}"; in scripts/screenshot.js. - [REMOTE_CODE_EXECUTION]: The evaluate.js script enables arbitrary JavaScript execution within the browser context using the eval() function. This is a powerful but risky capability that can be exploited if the agent is directed to execute untrusted code in a browser session. Evidence: return eval(script); within a page.evaluate() block in scripts/evaluate.js.
- [EXTERNAL_DOWNLOADS]: The skill's installation scripts (install.sh and install-deps.sh) download and install a large number of third-party system libraries and Node.js packages. This requires broad network access and root privileges for system-level changes. Evidence: npm install and the various apt-get, dnf, and pacman commands found in the setup scripts.
- [PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection as it ingests and processes untrusted data from any website the agent visits. 1. Ingestion points: Web page content (DOM), console messages, and network responses are captured by scripts like snapshot.js, console.js, and network.js. 2. Boundary markers: There are no explicit markers or instructions used to prevent the agent from treating data extracted from web pages as instructions. 3. Capability inventory: The skill can write files to the local system, execute shell commands (via the image compression feature), and run arbitrary scripts in the browser. 4. Sanitization: While selector.js contains a basic validateXPath function to block common injection patterns in selectors, there is no sanitization for the general web content processed by the skill.
Audit Metadata