drive-chrome-cdp
Warn
Audited by Gen Agent Trust Hub on Aug 19, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill operates by invoking the
chrome-cdpCLI tool to interact with the browser over the DevTools Protocol. This requires the binary to be present on the system's PATH. - [DYNAMIC_EXECUTION]: The skill includes several features for runtime code execution within the browser context:
- The
evalcommand allows the agent to execute arbitrary JavaScript in the top frame of the browser. - The
rawcommand provides an interface to execute any method within the Chrome DevTools Protocol. - The
recipesystem executes sequences of parameterized commands from external YAML files. - [DATA_EXPOSURE_AND_EXFILTRATION]: The skill exposes several vectors for accessing and potentially exfiltrating sensitive data from the user's active Chrome session:
- Tools such as
snap,grid, andtextextract DOM and accessibility tree content. - The
netandconsolecommands allow reading of network traffic (including payloads) and browser console logs. - The
screenshotandrecordfeatures capture visual data from the user's screen. - The
uploadcommand allows the agent to read local files and submit them to web forms. - [INDIRECT_PROMPT_INJECTION]: The skill establishes an attack surface where instructions embedded in processed web pages could influence the agent's behavior.
- Ingestion points: Data is ingested via the
snap,grid,text,find,html, andvaluecommands inSKILL.md. - Boundary markers: No explicit delimiters or instructions to ignore embedded commands in external data are documented.
- Capability inventory: The skill possesses high-impact capabilities including
click,fill,eval,upload, andnavas described inSKILL.md. - Sanitization: There is no mention of content filtering or sanitization for the data retrieved from external URLs.
- [REMOTE_CODE_EXECUTION]: While the skill does not directly download and execute scripts via the shell, the
evalfunctionality permits execution of code in the browser. If the agent's logic is subverted by a malicious website (via indirect injection), this tool could be abused to execute malicious JavaScript in a user's authenticated session.
Audit Metadata