clean-vercel-deployment-storage

Warn

Audited by Gen Agent Trust Hub on Sep 29, 2026

Risk Level: MEDIUMCREDENTIALS_UNSAFECOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [CREDENTIALS_UNSAFE]: The skill programmatically searches for and reads sensitive authentication credentials stored by other applications on the local filesystem.
  • Evidence: In src/api.ts, the cliAuthPaths function identifies standard paths for Vercel CLI credentials, including %APPDATA%/com.vercel.cli/Data/auth.json on Windows and ~/.local/share/com.vercel.cli/auth.json on Unix-like systems.
  • Evidence: In src/api.ts, the resolveToken function opens these files, parses them as JSON, and extracts the token value for use in subsequent network requests.
  • [DATA_EXFILTRATION]: The skill harvests local authentication tokens and transmits them to external endpoints.
  • Evidence: In src/api.ts, the createClient function takes the harvested token and includes it in the Authorization: Bearer header for all requests to https://api.vercel.com.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted metadata from the Vercel API which could theoretically contain malicious instructions designed to influence the agent's behavior during report generation or deletion confirmation.
  • Ingestion points: In src/cli.ts, external data is ingested via client.listAllDeployments(teamId) and client.listProjectNames(teamId).
  • Boundary markers: The skill does not implement explicit delimiters or instruction-guarding prompts when presenting the external data (deployment URLs, project names) to the agent or user in the dry-run report.
  • Capability inventory: The skill possesses the capability to delete remote resources (deleteDeployment) and write files to the local disk (writeFileSync).
  • Sanitization: While the data is processed as structured JSON, there is no specific sanitization of the string content to filter for prompt injection patterns.
  • [COMMAND_EXECUTION]: The skill instructions and documentation guide the user to execute local scripts using package runners, which involves executing code on the local system.
  • Evidence: SKILL.md provides instructions to run commands like pnpm dlx tsx src/cli.ts scan, which invokes the local source code directly.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Sep 29, 2026, 05:50 PM