Overview
MaelGard is a local checkpoint between your devices and generative AI services. The agent inspects outbound requests before they are sent, so the decision stays with you.
MaelGard combines a device agent, a console server and a signed catalogue of AI services. The agent inspects requests at send time. The console only receives metadata. Prompts and files stay on the device.
What MaelGard does
The agent redirects outbound HTTPS traffic to a local proxy. It only decrypts flows whose domain is listed in the catalogue and whose channel is active. The rest is relayed byte for byte, without inspection.
- On-device analysis : prompt text and attached files are examined locally by the embedded detection library.
- Checkpoint before sending : every request to an AI service in the catalogue goes through the local proxy, which applies the policy.
- Logged decision : the policy can allow, mask, ask the user or block. Every outcome produces a signed event.
- Metadata only : the console receives neither prompts nor files. It aggregates detected categories, services, decisions, timestamps and device pseudonyms.
Three-tier architecture
Three components cooperate: the device agent, the MaelGard server and the signed catalogue published by the vendor. Agent and server speak mutual TLS. Non-AI traffic is never decrypted.
| Tier | Role | Data handled |
|---|---|---|
| Device agent | Local TLS interception, detection, policy enforcement, decision window. | AI request text, detection models, signed local policy. |
| MaelGard server | Central policy, AI service inventory, dashboards, audit log, fleet management. | Event metadata, signed policy, device inventory, local catalogue. |
| Vendor catalogue | Signed list of AI services and their extraction rules, published by the vendor. | Ed25519 signatures, parsing rules, minimum agent versions. |
A request, step by step
- An application opens an outbound connection. The WFP driver or nftables redirects the TCP stream to the local proxy.
- The proxy reads the SNI. If the domain is not in the catalogue, the stream is relayed encrypted without inspection.
- If the domain is active, the proxy terminates TLS with a local certificate and reads the HTTP request.
- The catalogue extraction rule isolates prompt fields and attached files.
- The detection library returns the entities found, with category and score.
- The policy engine computes the outcome: Allowed, Masked, User decides or Blocked.
- The request is sent, rewritten or rejected. A signed event is logged.
Supported systems
| Platform | Mode | Network interception |
|---|---|---|
| Windows 11 x64 | MSI, Group Policy or Intune | WFP driver + local TLS proxy |
| Windows 11 ARM64 | MSI, Group Policy or Intune | WFP driver + local TLS proxy |
| Linux desktop x86_64 | Flatpak + host service | nftables + local TLS proxy |
| macOS | Not shipped | Planned, no committed date |
| Linux ARM64 | Not shipped | Planned, no committed date |
Design principles
- Nothing leaves the device in clear : detected sensitive values stay in the agent service memory, except toward the decision window of the user concerned.
- Non-AI traffic preserved : flows that do not target an active AI service are never decrypted.
- Isolated failures : the failure of a UI component must not block non-AI traffic.
- Server-side metadata only : the server never receives prompt text or file contents.
Who it is for
MaelGard is aimed at security teams, DPOs, IT and device administrators. The offering covers SMBs from 10 users up to large accounts with tens of thousands of devices.
- Security teams : visibility over AI usage, channel-based rules, alert review.
- DPO : AI service inventory, exports for the record of processing, sealed audit log.
- IT and administrators : Group Policy or Intune deployment, fleet management, device enrolment.
- Auditors : read access to the log, offline seal verification.