Skip to content

Security: unit-mole/reliabilityops-agentic-rag

Security

SECURITY.md

Security

ReliabilityOps is a portfolio/demo system, not a safety-certified industrial control system.

Implemented controls

  • Pydantic request/output schemas.
  • Query-length and tool-count limits.
  • Allowlisted analytical/database functions.
  • No arbitrary OS shell or Python execution tool exposed to the model.
  • No arbitrary LLM-generated SQL execution.
  • Parameterized/ORM database access with bounded results.
  • Retrieved documents are treated as untrusted data.
  • Indirect prompt-injection patterns are detected and evidence is marked untrusted.
  • Stable evidence IDs are checked during verification.
  • Lexical claim/evidence support checks supplement citation-existence checks.
  • Numerical finding values are checked for non-finite values.
  • Corrective actions require a separate human decision.
  • Commercial model keys are blank/disabled by default.

Secrets

Use .env; never commit credentials. .env is ignored by Git.

Production gaps

Authentication, RBAC, network isolation, secrets management, full SQL read-only roles, model sandboxing, rate limiting, and enterprise audit retention are future production work.

There aren't any published security advisories