dev-browser
Fail
Audited by Gen Agent Trust Hub on Feb 16, 2026
Risk Level: HIGHCREDENTIALS_UNSAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
- COMMAND_EXECUTION (HIGH): The skill makes extensive use of
execSyncinscripts/start-server.tsto manage processes and install software. More critically,SKILL.mdinstructs the agent to execute arbitrary TypeScript code by piping heredocs intonpx tsx, which bypasses standard script constraints. - CREDENTIALS_UNSAFE (HIGH): The 'Extension Mode' described in
SKILL.mdand the scraping guide inreferences/scraping.mdexplicitly target the user's authenticated browser sessions. The skill provides instructions for capturing and replaying authentication headers (e.g., cookies, bearer tokens) from intercepted network requests. - DATA_EXFILTRATION (HIGH): By connecting to a user's active browser session and providing methods to capture auth headers and page content, the skill facilitates the exfiltration of sensitive user data to external endpoints under the guise of 'scraping'.
- EXTERNAL_DOWNLOADS (MEDIUM): The
server.shscript andscripts/start-server.tsperform runtime installations of Node.js packages and Playwright browser binaries. While these use standard registries, they occur automatically during setup without integrity verification. - INDIRECT_PROMPT_INJECTION (HIGH): The skill lacks boundaries when processing external web data. It ingests untrusted HTML and API responses which are then used to drive automated actions (clicking, typing, file writing), creating a high-risk vector for adversarial instructions embedded in web pages to take control of the agent's browser session.
- Ingestion points:
page.goto(),getAISnapshot(), andpage.evaluate()are used to pull content from arbitrary URLs into the agent's context. - Boundary markers: None identified; untrusted data is processed directly as source for subsequent actions.
- Capability inventory: File system access (
fs.writeFileSync,screenshot), network access (fetch,page.on('request')), and arbitrary shell command execution. - Sanitization: No evidence of sanitization or validation of the ingested web content before it influences agent decision-making.
Recommendations
- AI detected serious security threats
Audit Metadata