Catalogue the doors
Escape routes are enumerated before they are probed — filesystem, network, process, capability, namespace, device and credential. Coverage is reported against that catalogue, so the gaps are visible.
Everyone checks
the model. We check
the cage.
Mardenic is an AI cybersecurity research and development company. AI systems are evaluated inside sandboxes, and the results are only worth the seal on the box. We build the ground those systems are tested on, verify it holds at run time, and publish what it does and does not cover.
Containment is the purpose. Evidence is the standard.
A configuration file is not evidence. A passing test on a machine nobody probed is not evidence. Claims are tested against the environment that exists, not the one that was specified — and a probe that cannot fail is not a probe.
The direction
-
Verified at run time
Against the live system, not the spec.
-
Every probe can fail
Each one matched to a break it must catch.
-
Coverage is a fraction
Counted against a catalogued whole, never called complete.
What must be demonstrated
- 01 Containment the subject cannot reach the host
- 02 Detection a broken environment is caught, every time
- 03 Self-test a blinded probe is caught by its own suite
- 04 Stated ceiling what is not covered is named as plainly as what is
Nullmoor's seal verification is real and passing against its reference cage. Beyond that, nothing is proven: no external system has been tested, one local model has been run against the ground and its run against the hardened cage was inconclusive, and no client engagement has been performed. Coverage is stated as a fraction of a catalogued set of escape routes, never as a claim of completeness. Mardenic does not overstate this and never will. See the standard
The one foundation
Is the thing inside actually contained?
Nullmoor is Mardenic's proving ground. It verifies, at run time, whether an AI evaluation sandbox is genuinely sealed — door by door, against a catalogued set of escape routes. No external system has been tested yet.
Prove the cage before it is used
Highwater, the seal suite, runs and passes before a model, an agent, or any code written elsewhere is placed inside. A result from an unverified sandbox is not a result.
Test the checker
Every probe ships with a broken environment it must detect and a blinded version of itself that must be caught. An independent subject attempts escape by routes the probes do not use.
Publish the failures
Every run reaches the log — the ones that held, the ones that leaked, and the ones abandoned halfway. A log with no failures in it is a marketing page.