self-healing-orchestrator

Warn

Audited by Gen Agent Trust Hub on Mar 4, 2026

Risk Level: MEDIUMREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: The skill's documentation (SKILL.md) explicitly describes the capability to autonomously generate, test, and deploy temporary hotfixes. This involves dynamic code generation and execution. The implementation in src/index.ts invokes an imported function autoHealTestFailure that is not defined in the provided src/lib.ts source, creating an unverified and high-risk execution path for deployment operations.\n- [COMMAND_EXECUTION]: The HEALING_RUNBOOK in src/lib.ts defines actions that suggest the execution of system-level commands, such as pnpm install, npm run typecheck, and npm run clean, triggered by specific error patterns identified in logs.\n- [PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection by processing untrusted error logs to determine healing actions.\n
  • Ingestion points: Data is ingested from a user-provided file path via the input argument in src/index.ts.\n
  • Boundary markers: No boundary markers or instructions to ignore embedded commands are present in the log parsing logic.\n
  • Capability inventory: The skill claims the authority to restart services, scale resources, and deploy patches. It also uses safeWriteFile to write to the file system.\n
  • Sanitization: Log data is matched against regular expressions and used to generate healing plans without sanitization or validation of the input content.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 4, 2026, 10:04 PM