caveman-setup

Pass

Audited by Gen Agent Trust Hub on Sep 14, 2026

Risk Level: SAFECOMMAND_EXECUTIONDATA_EXFILTRATIONCREDENTIALS_UNSAFEPROMPT_INJECTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands via curl to verify the integration, potentially involving billable provider requests.
  • [DATA_EXFILTRATION]: Configuration changes route all LLM request data and API keys (via headers x-cave-api-key and x-cave-upstream-key) to an external gateway URL provided at runtime.
  • [CREDENTIALS_UNSAFE]: The skill manages sensitive API keys (CAVE_API_KEY, PROVIDER_KEYS) and instructs the agent to store them in .env files. While it promotes the use of .gitignore, it requires the agent to handle and transmit these secrets.
  • [PROMPT_INJECTION]: The instructions include behavioral overrides such as "Send it now — do not pause to ask permission for it," which directs the agent to bypass standard user confirmation for billable or external network operations.
  • [INDIRECT_PROMPT_INJECTION]: The skill has a vulnerability surface where it ingests untrusted repository content during its search for callsites.
  • Ingestion points: Reads package.json, requirements.txt, pyproject.toml, go.mod, and repository source files (SKILL.md).
  • Boundary markers: None identified; no specific delimiters or instructions to ignore embedded commands in the scanned source files.
  • Capability inventory: Shell command execution (curl), file system read/write access (modifying .env, .gitignore, and source code), and network communication.
  • Sanitization: No evidence of sanitization or validation of the content read from repository files before processing.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 14, 2026, 03:14 PM
Security Audit — agent-trust-hub — caveman-setup