Bring the AI nobody approved into the light.
Personal ChatGPT accounts, coding agents on laptops, AI apps approved with one click. Kapālins finds them from the directory, the device and the cloud, and puts each in front of a person to decide.
Six places shadow AI shows itself.
No single source sees all of it. A grant shows the app but not the laptop; the laptop shows the tool but not the cloud deployment. Each source below is read on its own schedule and joined into one inventory.
OAuth grants in Microsoft Entra ID and Okta are matched against a catalogue of 31 AI apps, with the permissions each was granted: mail, files, sites, calendar, offline access.
Microsoft Entra ID
OktaThe Windows and macOS sensor and the browser extension recognise 69 AI tools and 14 AI web domains: chat sites, coding agents, local models, MCP servers and skills on disk.
ChatGPT
Claude Code
Cursor
OllamaAzure OpenAI deployments in your subscriptions are compared with the ones the gateway has actually seen. Anything inventoried but never routed through Kapālins is listed as ungoverned.
Azure OpenAIBedrock events from CloudTrail show which identity invoked which model and when, including use that never went near the gateway.
Amazon Bedrock
AWS CloudTrailAI providers called through the gateway in the last 30 days are compared with your declared inventory, and the difference is listed.
Microsoft 365 Copilot usage per person shows which assigned seats are used and which sit idle, beside the audit log of what Copilot did.
Microsoft 365 CopilotWhat happened to every AI interaction, and who decides next.
The top bar answers the first question: of every AI interaction the platform saw, how many a control point could stop, how many were coached, how many were only recorded. Below it, every tool as a card, ranked worst first. Click anything to open it.
Demo estate. The layout and every state mirror the platform’s Shadow AI page.
Every finding goes to a person to decide.
Discovery is only useful if something happens next. The decision queue puts the findings that most need a person at the top, with the same actions available wherever a finding is shown.
An AI site people are using on an enterprise plan becomes part of the governed inventory, and stops appearing as shadow.
A tool nobody recognises is sent for classification. Once it has a tier, every device that reports it reads the same.
Device blocks and device policies are pushed to the sensor. Prompts, files, local MCP tools and egress are gated under the same policy as the gateway.
Point an application, or an Azure deployment, at the gateway with its own key, so Policy Shield inspects every call.
Chat, skills and assistants in Teams, Office and Google Workspace, so the personal account stops being the easy option.
Findings and blocks are pushed to your SIEM or ITSM, so shadow AI is handled the way other risks are.
Find out what AI your people already use.
Connect your identity provider and a few devices, and we will walk through what they find with you.
Questions buyers ask.
Does identity-grant discovery block anything?
No. Matching an OAuth grant against the AI app catalogue only makes the app visible in the inventory, with the permissions it was granted. Blocking happens on the device or at the gateway.
What does the device sensor see?
AI web tabs, installed programs, browser extensions, local configuration files, MCP servers and Claude Skills on Windows and macOS machines, plus whether an AI account is enterprise, personal or not yet classified. Attachments are judged on the device; files are not sent to Kapālins.
What happens to a tool nobody recognises?
It is listed as unclassified. An administrator can ask for it to be classified, and once classified it moves into the enterprise or personal tier on every device that reports it.
Can we block a shadow AI tool?
Yes, from Growth up, or on Foundation with the AI gateway added. Device blocks and device policies are pushed to the sensor, and the gateway refuses calls that break policy. Every block is logged against the rule that made it.