Umbra
← All docs

AI Explorer · Tier 2 scenario playbooks

Ten scripted scenarios for the classes static rules can't reach (IDOR, OAuth bypass, JWT confusion, GraphQL abuse, mass assignment, price manipulation, role tampering, workflow bypass, multi-step ATO, cross-tenant UUID).

Last updated Jun 9, 2026

What it is

Tier 2 takes each Tier 1 candidate and runs it through a scripted scenario playbook, a bounded sequence of probes designed to confirm or refute the vulnerability class for that endpoint.

Each scenario is a small program with:

  • A fixed objective (“confirm IDOR on this endpoint”)
  • An HTTP probe tool the LLM can call (with the egress lock: only this scan’s target host is reachable)
  • A success criterion (“response leaks another user’s record” with concrete bytes)
  • A hard turn cap so the LLM can’t burn budget on a stuck case

What playbooks exist today

ClassProbes for
IDOR / BFLAObject IDs in URL or body that another tenant’s session can read
OAuth redirect_uri bypassOpen redirect or token-leak via redirect_uri manipulation
JWT algorithm confusionalg=none, HS256 vs RS256 confusion, key confusion
GraphQL abuseIntrospection-on, batching, alias overload, depth limit bypass
Mass assignmentHidden field assignment (admin=true, org_id=other) on user-controlled JSON
Price manipulationCart/checkout flows accepting negative quantities or external prices
Role escalationProfile-update endpoints accepting role/is_admin/permissions
Workflow bypassOrder-of-operations checks (skip approval, skip payment)
Multi-step ATOSignup → password reset → token leak chains
Cross-tenant UUIDUUID-based resources reachable across tenant boundaries

What it produces

For each candidate Tier 2 examines, one of three outcomes:

OutcomeWhat lands in the report
ConfirmedReal finding with the exact probe sequence as PoC
RejectedDrop the Tier 1 hypothesis; remember so Tier 3 doesn’t re-probe
Hypothesis preservedTier 2 was inconclusive; finding lands as “AI-inferred, not actively probed” at lower confidence

Destructive ops gate

Five playbooks (mass assignment, price manipulation, role escalation, workflow bypass, multi-step ATO) require the destructive-tests toggle ON, because they write state. With the toggle OFF the scan stays strictly read-only and these scenarios skip cleanly.

Why it matters

These ten classes account for the majority of real bugs in modern web apps, and they’re exactly what a static DAST scanner cannot find. Tier 2 is the wedge between “scanner that finds OWASP A03” and “audit that finds the bugs that get paid in bug bounty.”