A console for the whole picture.
Everything below works in the interactive demo (sample data), which runs on a tested engine (64 tests). The demo company is Kestrel Mutual, a made-up insurer with 2,400 staff and 30 days of generated activity.
How much of your AI use you can actually see
The Overview starts with the share of AI activity Umbra can see and classify. Next to it are the share of work on approved tools, how much sensitive data was contained, how many incidents are open, and a short weekly summary from Corona (AI assistant) with the most useful next steps. In the middle is Umbra at the centre: every AI tool at Kestrel Mutual on orbit, and each ring around a tool is a control that applies to it. Approved tools sit inside all six rings, unapproved ones further out, high-risk ones near the edge.
Every AI service on one map
Distance from the centre shows risk, size shows usage, and colour shows whether it is approved. Risk is computed from facts a buyer can verify: whether the vendor trains on inputs, how long it keeps them, which attestations it holds and where it processes data. Corona can also recognise AI services that are not in the catalog, just from the shape of the traffic.
Each prompt checked, with the reasons shown
The detectors check that a match is real before flagging it: card numbers must pass the Luhn check, bank account numbers the IBAN checksum, and US social security numbers the issuing rules. Each decision shows what was actually sent and which rules applied.
- Secrets, health data, personal and financial data, source code, codenames, confidentiality markers
- Stable placeholders, so a redacted prompt still makes sense to the model
Agents with owners and limits
The graph shows which people use which agents, and which agents call other agents and tools. Calls outside an agent's allowed list show in dashed red. It also shows indirect access: in the demo, the broker assistant can reach financial data it was never given, because it is allowed to call an agent that has it.
- Unowned agent detection: pause it, or approve it with an owner and limits
- Prompt injection and exfiltration scanning for agent payloads
The layer under the agents, with a pause that reaches Outpost
Models lists the AI models Kestrel's agents run on, with the same verifiable risk facts as the AI map, the data seen in each model's agent traffic, and the agents that depend on it. A person can pause an approved model, and every agent on it stops at once, in Umbra and in Outpost, where the bots fly home and their work goes back to the Inbox for the rest of the crew. Resuming the model frees exactly those bots. It answers the question a CIO asks the day a vendor has an incident: which of our agents break if we pull this model?
- Only approved models can carry agents; unapproved ones are governed by the rules
- Umbra lists its own drafting model here too, with its real test results
Write a rule in plain English, then test it on last month
Corona turns a sentence into a rule, points out anything it was unsure about, and restates the rule in English so a reviewer can check it. The draft comes from a real model (gpt-oss-120b on Groq), called through a small server of mine that holds the key; a checker drops anything outside Umbra's rule language, and if the model is down, Umbra drafts with its own rules-based parser and says so. Before the rule is turned on, "Test on last month" re-runs the last 30 days against it and shows how many decisions would change, and for which people and teams.
Fewer alerts, clearer decisions
Umbra groups related events into one case per decision someone has to make. Cases the rules already handled are closed automatically, with the evidence kept. Corona writes a short summary for the analyst and a friendlier note for the employee involved.
Where your company wants to use AI
Unapproved usage is grouped by team and task, then scored by how often it happens, how much of it runs on unapproved tools, and whether it is growing. The result is a now, next and later list of where to roll out approved AI. This is the part that turns a security tool into a planning tool.
Audit evidence on demand
Eight controls are measured from what the product is actually doing and mapped to the EU AI Act, ISO/IEC 42001, the NIST AI RMF and SOC 2. You can export an evidence pack, or a full inventory of AI models, services and agents in the CycloneDX format. A five-level maturity scale gives the board a simple way to track progress.
What employees actually see
Sensitive details are masked before sending, there is a one-click switch to the approved tool, and when nothing is wrong Umbra stays out of the way. The check runs on the person's own device before anything is sent.
