chrome-cdp
Pass
Audited by Gen Agent Trust Hub on Oct 7, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests data from external websites through commands like
snap(accessibility tree) andhtml(page source). This creates an attack surface where a malicious website could embed instructions to trick the agent.\n - Ingestion points: Commands
snap,html, andevalread data from the browser tab and return it to the agent.\n - Boundary markers: None identified in the provided scripts or instructions to separate site content from agent instructions.\n
- Capability inventory: The skill can interact with the browser via
click,type,nav,eval, andshot(screenshot) commands, allowing an injection to potentially trigger actions.\n - Sanitization: The script uses
JSON.stringifyfor selector interpolation but does not sanitize the content read from the page before returning it to the agent.\n- [DYNAMIC_EXECUTION]: The skill uses the Chrome DevTools Protocol'sRuntime.evaluatemethod to execute arbitrary JavaScript within the context of an open browser tab. This is the intended primary function of the tool but represents a high-privilege capability within the user's browser session.\n- [COMMAND_EXECUTION]: The script usesBun.spawnto launch background daemon processes for each browser tab. These processes manage the WebSocket connections and communicate via local Unix sockets located in the temporary directory. The target IDs for these processes are resolved from a list of active tabs provided by the browser.
Audit Metadata