AI Explorer · Tier 3 open hunt

What it is

Tier 3 is the final stage of the Deep Audit AI Explorer. It runs after Tier 1 + Tier 2 have generated, probed, and resolved their candidate list. Tier 3’s job is open-ended hunting: “forget the candidate list, given the full recon picture, what did we miss?”

It runs on Claude Opus (the heaviest reasoning tier) and has a longer turn budget than Tier 2. The context window includes:

• Clustered endpoint inventory (up to 200 representative clusters)
• JS-derived endpoint pointers (paths extracted from JavaScript bundles JupiterSec ran through LinkFinder)
• Discovered parameter names (from JupiterSec’s hidden-param fuzz)
• A SKIP list of endpoints Tier 1+2 already confirmed
• A “don’t retry without new evidence” list of Tier 2 rejections

What it produces

Open-form findings: anything Tier 3 can prove against the live target. Same evidence requirements as Tier 2: real probe sequence, real response bytes, confidence score.

Typical Tier 3 yield: 1-4 findings per scan, in the harder-to-find classes (chained auth bypasses, business logic, novel parameter abuse, encoded injection that the static modules’ payload pool didn’t try).

Why it matters

Scripted scenarios miss anything they don’t have a playbook for. Tier 3 is the catch-all that turns “scanner + 10 scenarios” into “agent that read the whole app and hunted.” It’s also where the ~30% of paid-tier audits surface findings the buyer didn’t expect, the wedge that justifies the $100 over a free scanner.

Cost model

Tier	Typical cost	Why
Tier 1	$0.20-$0.40	One Sonnet call over compressed inventory
Tier 2	$1.50-$3.00	Up to 30 candidates × Sonnet scenario calls
Tier 3	$0.80-$2.50	One long Opus session with bigger context

All three tiers + the Triager share the $30 hard cap per scan. Tier 3 fires only if budget remains.

Tunable

Disabled per-deploy via UMBRA_DEEPAUDIT_TIER3_ENABLED=false (for example, while bringing up a new environment or during a cost- tuning experiment). On by default in production.