Cyber Defense Harness¶
A platform that runs an organisation's own detection and response tradecraft as protected data, driving AI agents that investigate and respond across both IT and operational technology — power, cooling, building controls.
The platform is reusable and belongs to the vendor. The tradecraft is yours, loaded as structured data the platform reads at run time, never compiled into it, and never visible to another organisation. That split is the product.
What makes it different¶
-
Tradecraft is data, behind a hard boundary
Your rules load as declarative playbooks. Change a threshold and the next run behaves differently — no rebuild, no restart, no deployment. An automated check fails the build if any customer content reaches platform code.
-
Tiered autonomy, with a hard OT gate
Observe → recommend → act-with-approval → act alone. Anything that changes state waits for a person until that action type has earned trust through a measured record. Operational technology is capped at recommend, always.
-
Model-agnostic, and runs disconnected
Frontier, open-weight or local inference. Air-gap mode removes every route off the host — there is no egress to block rather than egress that is blocked.
-
One platform, many organisations
The same alert produces different verdicts for different customers, from the same binary, with neither able to see the other.
Where to start¶
| If you are | Read |
|---|---|
| New to any of this | Glossary — the vocabulary, in plain language |
| Standing the system up | Installing |
| Responsible for what agents may do | Operating the harness |
| Reviewing what the agents found | Using the harness |
| Writing the rules it runs | Authoring tradecraft |
| Showing it to someone | The demonstration, narrated |
In one minute¶
uv sync
./deploy/bin/harness up # console on :8080, API on :8000
./deploy/bin/harness demo # the ten-minute script, headless
The demonstration drives the public API only and never touches the browser. The console is a client of that same API — there is no endpoint reserved for it, and a test asserts so. Anything the console does, a script can do.
Status
This is a proof of concept. All eleven acceptance criteria pass — proven by tests that run, not by files that exist — and the demonstration completes three times from a clean reset.
It is still not a production security platform. There is no backup or retention, no high availability, no monitoring, no connectors to real products, and the data is synthetic throughout. Identity is real (Keycloak, with OIDC) and a deployed instance gets real TLS, but there is no user management, no MFA and no brokering to a customer's own directory. Each document states what works today and marks what does not.