agent-first-screenshots
Warn
Audited by Gen Agent Trust Hub on Aug 14, 2026
Risk Level: MEDIUMREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The
loadSharpfunction in bothscripts/capture-verify.mjsandscripts/screenshot-verify.mjsdynamically resolves thesharplibrary by searching the filesystem for specific.pnpmdirectory structures and importing from computed paths at runtime. - [REMOTE_CODE_EXECUTION]: The
scripts/capture-verify.mjsscript accepts apreCaptureparameter which is executed as an arbitrary JavaScript string in the browser context via theRuntime.evaluatemethod of the Chrome DevTools Protocol (CDP). - [DATA_EXFILTRATION]:
scripts/capture-verify.mjsinitiates a WebSocket connection to an arbitrarywsUrlprovided as a command-line argument, which could be used to facilitate communication with untrusted endpoints. - [PROMPT_INJECTION]: The
preCapturefeature creates a surface for indirect prompt injection, as an agent could be manipulated by the content of a target application to execute malicious JavaScript in the browser session. - [COMMAND_EXECUTION]: The skill's operation depends on the execution of Node.js scripts that manage browser state, capture screenshots, and perform file system writes to specified and temporary paths.
Audit Metadata