NEW RESEARCH: Your Sandbox Is Made of Glass

Read

Trinitite

PricingResearchBlogPodcasts

Shadow AI Inventory

“How do you know your AI inventory is complete?”

You blocked OpenAI at the firewall — then a confidential model showed up in prompt logs from a free chatbot anyway. Trinitite consumes the Secure Web Gateway telemetry you already have, correlates it against your governed traffic, and emits a signed, deduplicated inventory of unsanctioned AI usage. No new inline blocker to defend.

shadow_ai · summary · 30d

SIGNED

342

shadow events · 41 users

chat.openai.com

198 · shadow

*.anthropic.com

76 · shadow

*.cursor.com

54 · governed

sanctioned

1,204

read-only

no block injected

We never inject into your SWG. We only read what it already exports.

The platform contributes the curated registry of AI endpoints, the correlation against governed traffic, and the signed evidence bundle the auditor can cite — leaving block decisions where your security team already trusts them. A senior partner can paste an M&A model into a free chatbot in twelve seconds; detection is the first defensible step.

How it works

Read, classify, correlate, flag, sign.

01

Read SWG telemetry

Consume the DNS and URL events your Secure Web Gateway already exports — we never inject, decrypt TLS, or see request bodies.

02

Classify against the registry

Match each event against ai_destination_registry — Trinitite’s curated, maintained catalogue of known AI vendor endpoints, chat surfaces, and SDK CDNs.

03

Correlate ±60s

Check for a governed request to the same vendor by the same principal within ±60 seconds — the sanctioned Trinitite path counts as evidence the usage was governed.

04

Flag shadow

A known AI vendor with no governed equivalent in the window is flagged is_shadow, deduplicated to a counted single row per 24 hours.

05

Sign the inventory

Events are append-only and chain-linked into the unified ledger; the summary emits signed by_vendor, by_user, and totals rollups ready for a workpaper.

Zscaler NSS

Cisco Umbrella

Cloudflare Zero Trust

Netskope

Forcepoint

generic DNS

The signed bundle

Three rollups, ready for a workpaper.

by_vendor

vendor, category, event_count, shadow_count, user_count — answers "where is the unsanctioned usage going?"

by_user

matched_user_id, raw principal, event_count, shadow_count — answers "who is doing it?" resolved to your IdP where possible.

totals

events, shadow, sanctioned, unknown_destinations — the top-line numbers for the executive summary and trend chart.

In your language

A complete inventory, in your terms.

CISO

A signed, queryable inventory of every unsanctioned AI vendor visit by user, device, and time — from telemetry you already pay for, with no new inline blocker to defend in procurement.

Chief Audit Executive

"How do we know our AI inventory is complete?" — answered by a continuous, deduplicated, externally clock-anchored stream instead of an annual survey.

Compliance Officer

EU AI Act Annex IV §1 inventory, SR 11-7 §III shadow-model inventory, SOC 2 CC6.6, SOX §404 — all backed by the same signed event stream.

General Counsel

GDPR Art. 28: unsanctioned AI vendors are uncontracted processors. Detection is the first defensible step toward DPA closure or vendor enforcement.

By design

What it deliberately does not do.

Inject blocks into your SWG — block decisions stay where your security team already enforces them.

Decrypt TLS — we consume domain / URL telemetry the SWG already extracted; never request bodies.

Mass-store raw logs — unmatched, unflagged events are dropped to keep the inventory consumable.

The governed-route content controls live in reversible masking; the audit substrate is the audit platform; the privacy program is for CPOs & compliance.

FAQ

Shadow AI inventory, answered

What is shadow AI?

Shadow AI is unsanctioned AI use — employees pasting sensitive work into ChatGPT, Claude, Gemini, Perplexity, or Cursor outside your governance perimeter, with no log, DLP signature, procurement record, or DPA. Trinitite’s Shadow AI Inventory consumes the Secure Web Gateway telemetry you already have, correlates it against your governed traffic, and emits a signed, deduplicated inventory of that usage.

Is this another inline blocker I have to deploy?

No — and that is deliberate. Trinitite never injects into your SWG; it only reads what the SWG already exports. The platform contributes the curated registry of AI endpoints, the correlation against governed traffic, and the signed evidence bundle — leaving block decisions where your security team already trusts them. That one-way design is what makes it shippable without a procurement battle.

How do you avoid false positives?

A SWG event is flagged shadow only when both hold: the destination matches a known AI vendor in the registry, and the same principal has no Trinitite-governed request to the same vendor within ±60 seconds. The window distinguishes an ungoverned ChatGPT visit from a sanctioned ChatGPT-via-Trinitite visit — two events that look identical to the SWG but mean opposite things to the regulator. Identical events collapse to one counted row per 24 hours.

How does it fit the rest of the platform?

Shadow AI feeds the quarterly signed AI inventory attestation, the agentic risk score, vendor-confirmation outreach, CCT assertions, and embedded-insurance pricing — cleaning up shadow AI tightens your premium. For content-level visibility on the governed route, see reversible masking; for the audit substrate, the audit platform.

Turn the SWG telemetry you already have into a signed inventory.

Point us at a read-only export from Zscaler, Umbrella, Cloudflare, or Netskope — get back a deduplicated, signed inventory of unsanctioned AI usage by vendor, user, and time.