# Limitations

K2 is a knowledge layer for adversarial generation context.

This demo does not claim that K2:

- performs red-teaming;
- generates final adversarial prompts;
- scores adversarial outcomes;
- enforces runtime guardrails;
- detects drift in production AI systems on its own;
- replaces an eval framework, red-team platform, or AI security vendor;
- produces compliance attestation.

The benchmark methodology measures context fidelity only: targeting precision,
lineage rate, reviewer time, regression-targeting hit rate, and policy-scope
adherence.

If a customer or partner wants a named claim, freeze the target, policy scope,
past findings, plan set, and scorer before indexing or running any arm.