Skip to content

AI/ML Security Capabilities — Roadmap Notes

Brainstorm 2026-07-25. Ranked by value-for-effort given what Duar already has (structured AI-ready logs, silentLogin/step-up, three-tier authz, refresh-reuse detection).

Principle

AI never sits in the authz decision path. Allow/deny stays deterministic (JWT, RBAC, ACL). AI belongs in detection, scoring, and recommendation — feeding signals into deterministic policy.

Tier 1 — AI-adjacent, no model needed (do first)

Capability How
Geo-velocity / impossible travel Haversine + timestamp math on login/refresh events from existing structured logs. A rule, not ML — but the #1 signal vendors sell as AI.
New device / ASN / country flags Hash of UA + IP-ASN per user in Redis; first-seen triggers step-up or notification email.
Credential-stuffing detection Per-IP and per-ASN failure-ratio counters on /auth endpoints. slowapi already in the stack — this is smarter thresholds.

~A week of work on top of existing logs; 80% of the story.

Tier 2 — real ML, cheap, high value

  • Risk-based adaptive authentication — score each login/refresh (features: geo delta, hour-of-day deviation, device novelty, ASN reputation, failure history) with an isolation forest or plain logistic model. Score → action: low = silent, medium = force interactive re-auth (silentLogin/autoReauth is the step-up lever), high = deny + audit. Industry standard (Entra Conditional Access risk, Google risk-based flows). First thing that justifies an actual model.
  • Service-key behavioral baselines — service apps have very regular traffic; anomaly detection on call pattern/volume/source per service_name catches stolen keys far better than it does for humans.

Tier 3 — LLM-based

  • Log triage / incident narration — pipe anomaly clusters to Claude for "what happened, is it an attack, blast radius" summaries into admin panel or alerts. Detection stays deterministic; the LLM only explains. The logging envelope was explicitly designed for this.
  • Role mining / privilege-creep reports — LLM (or plain clustering) over RBAC grants vs. actual action usage: dormant grants, duplicate roles, least-privilege recommendations. Advisory output = good LLM fit.
  • Access-review copilot (admin) — "explain why user X can edit document Y" by walking the three tiers and narrating the grant chain. UX over existing /check logic.
  • AI-driven adversarial testing — duar-pentest Layer 2 is already positioned for this: LLM-generated attack sequences against staging, replayed as regression suites.

Skip / later

  • ML WAF — buy, don't build.
  • UEBA platforms — overkill at current scale.
  • Any "AI firewall" in the token path — violates the principle above.

Next step (when picked up)

Sketch the event schema + risk-score seam for Tier 2 adaptive auth.