AI validation

Umbra’s AI agent runs real exploitability checks against your services and writes one-page proofs you can hand to the team that needs to fix the issue. This page is what it does, what it costs, and how to keep spend predictable.

What the agent actually does

For each CVE you click Verify on, the agent:

1. Pulls every public CVE record + advisory + exploit-DB entry it can reach.
2. Builds a working hypothesis: “to exploit CVE-X against this service, send this kind of request and look for this kind of response.”
3. Issues a small number of non-destructive test requests against your service (no data modification, no auth abuse, no DoS, just the minimum signal needed to confirm exploitability).
4. Writes one of three verdicts:
   • exploitable: chain reproduced, here’s the proof
   • not exploitable: chain attempted, here’s why it failed (often: “service is patched”, “behind WAF”, “wrong version”)
   • inconclusive: couldn’t determine safely; manual review needed

The full transcript (every tool call, every response sniffed) is saved in the run record so you can audit what the agent did. Verdicts come with a markdown proof you can paste into Jira.

What it costs

AI runs are metered in real dollars. Your bill reflects the actual Anthropic API cost × a ~2× markup that funds our infrastructure + margin. Typical run costs:

• Per-CVE validation: $0.02 – $0.20 (most criticals run in this range; complex chains cost more)
• Full-agent run (run the agent against a service with no specific CVE, let it discover what it can): $0.50 – $3.00 typical
• Re-validation: same as the original verdict

Each paid plan includes a bundled monthly AI allowance:

Tier	Bundled AI / month
Starter	$20
Pro	$100
Business	$400
Enterprise	metered, volume-discounted

Bundled allowance resets at billing-cycle close (use-it-or-lose-it).

Top-ups

When the bundled allowance runs out, Umbra falls back to your top-up balance. Top-ups are fixed SKUs ($25 / $100 / $500) and persist with a 12-month expiry per batch. Buy once, drain over the year.

Drain order is always bundled → topup → over-cap. Over-cap runs are allowed but capped per-run (see ceiling below).

Per-run ceiling

Every run has a hard dollar ceiling (by default $5 per run). If the agent’s costs would exceed the ceiling, the run aborts with an “exceeded ceiling” verdict and Umbra eats the overage, never the customer. The ceiling is adjustable in Settings → AI (superadmin only on v1; per-org tunable later).

What the agent will not do

• Modify data: no PUT/POST/DELETE on real customer endpoints unless explicitly opted in per-target.
• Test credentials brute-force: no login attempts; that’s a separate nuclei default-creds workflow that lives outside the AI budget.
• Resource abuse: no DoS, no traffic flooding, no API quota burn. A run that requires that gets reported as “manual review needed” instead.

Disabling AI entirely

If your security policy forbids LLM-assisted testing, the agent can be disabled for the entire org from Settings → AI → Disable AI. The Verify button disappears from the Findings page; all other Umbra functionality continues to work.