Where PromptShields fits
The honest framing for a budget conversation: PromptShields is not another DLP product. It is the in-prompt layer that your existing DLP, CASB, and MDM stack structurally cannot observe.
Status at a glance
Identity
Microsoft Entra ID
Single sign-on for the admin console, with tenant mapping so everyone from one Azure directory lands in one workspace. Covered in full on Identity and access.Auth0
The endpoint clients authenticate through Auth0:Device management
Microsoft Intune
The connected MDM integration. PromptShields reads your managed-device roster through Microsoft Graph and cross-references it against client check-ins, so the console can tell you which machines are genuinely covered rather than merely targeted by a policy. Intune is also the delivery mechanism — force-installing the browser extension and deploying the desktop agents. See MDM rollout.Jamf Pro, Kandji, and JumpCloud
Roster adapters for these three are planned. They will follow the same shape as the Intune integration: read the device list from the provider’s API, normalise it to a common device record, and reconcile against client check-ins.The force-install policies work today on all of these platforms. The adapter only affects whether PromptShields can read your inventory back — it is not a prerequisite for deploying the clients. Payloads for each are on the MDM rollout page.
SIEM — Microsoft Sentinel
The integration uses a deliberate two-channel design, because the two things a SOC needs from this data are not the same thing.
Why both: alerts in Sentinel are first-class objects with an analyst queue, automation, and SLA tracking. If every event arrived as a plain log row, your SOC would have to write a detection rule before they saw anything at all. The separate alert channel gives them an inbox on day one.
The legacy Log Analytics HTTP Data Collector API is deprecated and is not used. Syslog/CEF forwarding is not appropriate for this data and is not offered.
Setup
1
Create the Azure resources
Run the provided Bicep template in Azure Cloud Shell. It creates the Data Collection Endpoint, the Data Collection Rule, the custom table, and the role assignment.
2
Register an app in your tenant
The app registration lives in your directory. Grant it
Monitoring Metrics Publisher scoped to the specific DCR — not tenant-wide.3
Connect it in PromptShields
Supply your Tenant ID, DCE URL, DCR Immutable ID, stream name, and table name in the admin setup wizard.
4
Verify with KQL
Authentication prefers a Federated Identity Credential over a stored client secret. FIC removes secret rotation entirely and eliminates the blast radius of a leaked secret. Use it unless your policy forbids it.
What lands in your workspace
UserAadObjectId is denormalised into the table so KQL joins against SigninLogs work without a lookup table — you can correlate AI activity with sign-in activity directly.
Delivery guarantees
Every event either lands in Sentinel or is visibly in the dead-letter queue — no silent loss. Idempotency is keyed onEventId.
A future ASIM parser will map the table to
imAuditEvent so existing ASIM-based queries work unchanged. A Content Hub connector, workbook, and analytic rules are planned to follow the event stream.
Data classification — Microsoft Purview
PromptShields ingests documents through a Purview adapter, feeding classification context into the risk register. This is complementary rather than overlapping: Purview classifies data at rest and in M365; PromptShields observes what gets typed into third-party AI tools. Where a metric in the console draws on more than one source, its provenance is shown — so a number reading “Source: Purview DLP + endpoint audit” can be clicked through and verified rather than taken on trust.Enterprise architecture — Ardoq AI Lens
For organisations running Ardoq, the console generates a nine-file CSV import bundle describing your AI estate:
Inventory metadata only — no prompt data is included.
Collaboration — Slack
A Slack bot collects AI use-case declarations directly from staff, which is considerably more effective than emailing a form and hoping.Consent is built in. The first DM identifies the bot, names the organisation, states that responses go to the IT lead only, and offers a permanent opt-out by replying
STOP. Opted-out users are skipped by the delivery worker from then on. Raise this proactively with your works council or employee representatives — it is the question they will ask.Product analytics
The endpoint clients can emit product-usage analytics through PostHog, Firebase, and Google Analytics. These are configured per environment and can be disabled at build time.Billing — Stripe
Subscription and billing management is handled through Stripe. The Windows agent exposes aprompt-shields://billing deep link that routes users to subscription management; the macOS agent and console link to the same flow.
This is relevant to IT only in one respect: if you restrict protocol-handler registration by policy, allow-list the prompt-shields:// scheme or both sign-in and billing links will fail. See Windows agent.