HEIMDALL

Tech Specs.

HEIMDALL is the firewall for AI: it checks every request to ChatGPT, Claude, Gemini & Co. before it leaves your house. It replaces sensitive details with placeholders, translates the answer back locally and logs every decision. It works by your own policies, on your hardware. So there is a sanctioned path to cloud AI, instead of shadow AI.

Two risks you cannot switch off.

Your people have been using cloud AI for a while now, officially or quietly. That means you take on two risks, and good intentions alone fix neither of them.

The provider.

Corporations negotiate their own contracts. What you get is the training toggle. Whether and how your inputs are stored and processed follows the provider’s terms. You cannot audit them.

The human factor.

Even trained employees slip: the customer name in the prompt, the contract excerpt in the attachment, the rush before a deadline. Training is still a must, but it won’t stop slips like these.

That is the uncertainty you work with today. Every day.

Replace instead of block.

HEIMDALL protects without forbidding anything. It replaces sensitive details with placeholders before the request leaves your house, and puts them back in the answer. Take a look at the four steps.

01You ask the way you always do.

You write real names and real figures. Nobody has to hold back or paraphrase to keep data in. So the protection fits into everyday work.

02HEIMDALL checks before anything goes out.

HEIMDALL detects what your policy marks as sensitive and shows the finding, the rule and the placeholder. Then you decide: send protected, handle locally or cancel.

03The cloud only sees placeholders.

⟨CUSTOMER_A⟩ instead of K-20388: the external model understands your question but never learns the real names and values. And it still answers at full quality.

04You read the real names.

The answer is translated back locally: your screen shows the original values, a discreet note says what was protected, and the audit log holds the evidence.

Using HEIMDALL costs you nothing in quality. That is why it gets used instead of bypassed.

Real names inside. Only placeholders outside.

The round trip of a request: inside the house it carries real names, the cloud receives only placeholders, and the answer is translated back locally.

Your houseThe cloudProtection
Write Karin Wolters a reply to her complaint about the press in Hall 2. Customer number K-20388. The credit goes to DE89 3704 0044 0532 0130 00. Write ⟨PERSON_A⟩ a reply to her complaint about the press in Hall 2. Customer number ⟨CUSTOMER_A⟩. The credit goes to ⟨IBAN_A⟩. 3 details protected. Policy “IT Policy §4.2: Cloud AI Use”.
Rewrite this clause in formal contract English: if downtime exceeds 48 hours, customer K-20417 pays a penalty of €42,000 under contract V-2019-088. Rewrite this clause in formal contract English: if downtime exceeds 48 hours, customer ⟨CUSTOMER_A⟩ pays a penalty of ⟨AMOUNT_A⟩ under contract ⟨CONTRACT_A⟩. 3 details protected. Policy “IT Policy §4.2: Cloud AI Use”.
Why does this call fail? curl -H "Authorization: Bearer sk-live-8f3kq92m4x" api.example.com/v1/orders Why does this call fail? curl -H "Authorization: Bearer ⟨SECRET_A⟩" api.example.com/v1/orders 1 secret protected. Built-in rule, no AI can overrule it.

That is all the provider sees. Translated back locally: you read the real values.

Examples. Names, numbers and keys are invented.

Six reasons this works in practice.

Live in one afternoon.

Upload the policy, confirm the rules, start observation mode. One appointment is all it takes.

Rules from your own documents.

You do not write rules. HEIMDALL compiles them from the policies you already have. You review and approve.

Hard rules for the known.

IBANs, credit cards, tax IDs, API keys and passwords: built in, and no AI can overrule them.

AI for the unexpected.

A local model catches what no list contains: the project name, the internal shorthand, the delicate connection.

Auditable evidence.

Every decision is in the audit log: what was detected, what left the house, who decided.

Provider neutral.

Works with any provider offering an OpenAI-compatible interface. Change the model and your protection stays.

Three sources. One policy.

HEIMDALL enforces the rules that already apply in your house.

1Built in

IBANs, credit cards, tax IDs, API keys, passwords: the baseline every company needs. These rules are immutable, and no AI can overrule them.

2Compiled from your documents

Upload your IT policy, NDA templates or works agreement. HEIMDALL proposes concrete rules from them, quoting the passage it found. You confirm, adjust or discard.

3Proposed from daily use

When the checker spots something sensitive that no rule covers, it proposes a new one. Your policy grows with everyday work. The last word stays with you.

When in doubt, HEIMDALL plays it safe. If the check is unsure, it asks you.

The policy editor in LOKI: the policy “IT Policy §4.2: Cloud AI Use” with its confidence threshold, the detected entity types such as person, location and organisation, and custom regex rules, for instance for project numbers.

Every rule stays visible and adjustable: entity types, custom patterns, confidence threshold. Per policy, per group.

Unobtrusive. But visible.

We refuse to run a covert filter: because of co-determination, because of transparency obligations, and because every employee should see daily that the company protects its data.

The notice in the tool.

When a cloud model is active, LOKI says so openly. And it says that HEIMDALL checks every outgoing request against company policy.

The compliance dialog.

In warning mode HEIMDALL asks whenever content is flagged, naming the policy that applies. So you always know why it asks.

The quiet default.

In regular operation “send protected” is the default: one discreet note per intervention, not one click more than necessary.

The notice above the input field in LOKI: “Cloud model active: your messages and uploaded documents are sent to an external provider. HEIMDALL is active and checks all outgoing data against your company policies.”

The notice sits right at the input field. So everyone sees it, every day.

One afternoon to running.

1Upload

Your existing policy, your templates. As PDF or Word, exactly as they are.

2Confirm

HEIMDALL proposes rules and you walk through them sentence by sentence: confirm, adjust, discard.

3Measure

HEIMDALL tests itself against your own policy and measures its accuracy before it enforces anything.

4Observe

In observation mode you see what HEIMDALL would have caught, without it intervening.

The rollout is built in. In three stages.

Observe, Warn, Enforce. Enforcement only starts once your metrics reach the agreed corridor.

We measure accuracy on your policy. And we do it before activation.

Inside LOKI. At the edge towards the outside.

HEIMDALL is built into LOKI and can also be booked on its own. It sits right where a request would leave your house. Local models never pass through it, because everything stays in the house. Only when you pick a cloud model in LOKI does it check every request on the way out. And on the way back the answer passes through it again: the placeholders are translated back into the real values, locally.

Your house

LOKI · Documents · Mails · Databases. Outbound: placeholders instead of real names.

The cloud

ChatGPT · Claude · Gemini · Mistral · …. Inbound: the answer, translated back locally.

Built for the audit.

Every outgoing call appears in the compliance audit log: what was detected, what left the house, how it was decided. That turns “we are careful” into evidence.

GDPR

Data minimisation and accountability: placeholders instead of personal data, and every decision documented.

Trade Secrets Act

Trade secrets are only protected if you can demonstrate appropriate measures. HEIMDALL is such a measure, log included.

EU AI Act

Operator duties and transparency: you know and can prove which data reached which AI system.

NIS2 · ISO 27001

Control over outbound data as a technical measure you can audit.

Features

And that is far from everything.

Policies

  • Manage policies in the dashboard
  • Ask, replace or block
  • For everyone or per group
  • Several policies, merged strictly
  • Pick categories with a tick
  • Confidence threshold per policy
  • Faulty patterns are rejected
  • Activate with one click
  • Managed by administrators only

Detection

  • People, companies, places via AI
  • Detection entirely in house
  • Email addresses, IBANs, credit cards
  • Checksum test for IBAN and card
  • Custom patterns such as project numbers
  • No AI overrules custom patterns
  • Texts in German and English
  • Document passages sent along, too
  • Text from attachments, too
  • Outlook mail content, too¹

When sending

  • Warning when a cloud model is selected
  • “HEIMDALL is active” at the input field
  • For every cloud model in LOKI
  • No finding, no prompt
  • Findings with category and value
  • Replace, handle locally or cancel
  • After “Handle locally” the chat stays local
  • Cancellation explained in the chat
  • Document passages replaced without a click
  • Value and placeholder in the dialog
  • The request as sent, to read back

On the way back

  • Answer translated back locally
  • Translated back as it streams in
  • Same value, same placeholder
  • Placeholders in the interface language
  • Altered brackets recognized, too
  • Unknown placeholders are never guessed
  • The history keeps the real values
  • Earlier answers protected again
  • List of replaced document content

Evidence

  • Every decision in the audit log
  • Who, when, which policy
  • Category and confidence per finding
  • Cancellations and local redirects, too
  • Search by user, tool, policy
  • Filter by date, decision, tool
  • Log for administrators only
  • The dialog notes the logging
  • Every cloud call with model and tokens

¹ With Microsoft 365.

HEIMDALL runs on your hardware, as part of LOKI. We only quote accuracy figures once they have been measured on your policy. And they are measured before activation.