Skip to content

Security: Git-Kapish/OpsPod

Security

docs/security.md

Security and data handling

Everything in this repository is synthetic: a fictional company, invented people, no real payments, no real PII. The controls below are built as if the data were real.

Provider data statement

The LLM providers are free tiers (OpenRouter :free models, Google AI Studio). Free-tier providers may log prompts and use them for training. This is the reason all data is synthetic and all PII is redacted before any LLM call. Do not point this system at real customer data on a free tier.

Controls (PRD S1-S10) and where they live

Control Implementation Test
S1 permissions Tool registry: each tool declares a scope; callers are Principals with scopes and caps (opspod/tools/registry.py) tests/integration/test_tools.py
S2 spending limits Per-case and per-day caps (tool layer, defense in depth), per-customer refund frequency (policy), all from config/policies/*.yaml test_tools.py, test_policy.py
S3 approvals Only approver/admin may decide; who/when/why recorded; edited amount can only reduce test_approvals.py
S4 idempotency UNIQUE(idempotency_key) on actions, key (case, kind, version); concurrent duplicates pay once test_ledger.py, test_approvals.py::concurrent
S5 reversibility Reversal is a new linked ledger entry; ledger and audit rows cannot be updated or deleted (Postgres triggers, also blocks TRUNCATE) test_ledger.py::test_append_only
S6 audit Append-only audit_log: input hash, context, prompt version, model, raw model output, proposal, rule ids, action, actor, trace id. GET /v1/cases/{id}/replay re-derives the decision from the record alone test_approvals.py::replay
S7 data minimization redact_pii removes names, emails, phones, addresses and card fragments before any LLM call; the mapping stays server-side; the sent fields are logged by category only test_skills.py, test_pipeline.py::test_llm_never_sees_pii
S8 prompt injection Detection before the LLM; case text wrapped as untrusted data; the model has no tool access and returns a schema-validated proposal; its own output is scanned for override instructions; the policy engine decides routes tests/chaos, test_eval_gate.py
S9 escalation Low confidence, conflicting evidence, fraud signals, repeat disputes or any validation failure routes to a human with a prepared summary test_pipeline.py
S10 fail safe Any exception, timeout, invalid output or unknown state routes to a human or the DLQ; never auto-execute tests/chaos

Trust boundaries

 untrusted                       |  semi-trusted            |  trusted
 customer complaint text         |  LLM output (proposal)   |  policy engine, ledger,
 webhook payloads                |  free-tier provider      |  tool registry, DB
 ---------------------------------------------------------------------------------
 detect_injection + redact_pii -> LLM -> schema validation -> policy engine -> tools

The model never sees raw PII and never holds a tool. Its output is data that the policy engine may reject. Money amounts are computed in code; the model only chooses among a fixed set of bases.

Secrets

API keys come from the environment only (OPENROUTER_API_KEY, GEMINI_API_KEY); .env is git-ignored; .env.example has blanks. Keys are held as SecretStr and sent in headers, never in URLs (the Gemini key goes in x-goog-api-key). Logs carry ids and hashes, not complaint text.

Known limitations (honest list)

  • Auth is a placeholder. API keys map to roles via the API_KEYS setting; the defaults are dev keys (dev-admin, ...). Replace before any shared deployment. There is no user management and no rate limiting on the API.
  • The console stores the key in a HttpOnly, SameSite=Strict cookie; it has no CSRF tokens beyond that, and no session expiry.
  • The console loads htmx from a CDN with a pinned version and an SRI hash.
  • POST /internal/alerts needs the alerts key; the default is a dev key.
  • Chaos injection is honoured only when APP_ENV=dev.
  • Context (trips, payments, history) comes from a fixture-backed repository, not production systems; the ContextRepo protocol is where a real integration plugs in.
  • Real-model numbers exist for one free model on a small set (see EVALS.md); prompts were sent to that provider (synthetic, PII-redacted). .env.local holds the key and is git-ignored.

There aren't any published security advisories