Agent With a Weapon
What it is
Agent or tool-execution paths are present without enough proof of sandboxing, scoped permissions, and prompt-injection boundaries.
This usually appears when model input, retrieval, or tool calls are added faster than capability boundaries and adversarial tests.
How to fix it
Add tool allowlists, capability scopes, sandboxed execution, replay logs, and prompt-injection tests.
- 1Define tool allowlists and capability scopes before tool execution.
- 2Sandbox execution and retain replay logs for sensitive actions.
- 3Add prompt-injection and untrusted-input boundary tests.
How VibeAudit checks for this
VibeAudit checks this with SAST_LOCAL_STRUCTURAL AI/tool route evidence and AI_CODING_AGENT_STATIC indicators for agent or IDE autorun files when those public or local files are in scope. The free tier may show the surface name only, without private evidence or remediation details. Metadata-only surface maps do not create missing-control findings. Unsupported private repositories, authenticated behavior, destructive workflows, account configuration, and business logic remain unverified unless the scan includes matching local, imported, or authorized runtime evidence.
Related failure modes
Check your public launch surface and keep unsupported scope visible.
Run Quick Scan