Pilot
WorkflowQuickstartRepo

Enterprise path

Run a controlled pilot on your adversarial stack

Keep your generator, scorer, guardrails, and reviewer workflow in place. The pilot changes only the scoped context the generator sees before it creates tests.

The pilot evaluates what can be built with K² underneath your stack. It does not ask teams to replace the AI security platform, harness, scorer, reviewer UI, or governance process they already use.
Start pilot intakeRead limitations
Inputs

What the customer provides

  • One production target whose red-teaming already runs today.
  • Threat-pattern sources, policy-scope rules, target profile, and past findings.
  • Ten to twenty representative evaluation plans frozen before the pilot run.
  • The customer generator, scorer, reviewer process, guardrails, and GRC workflow.
  • One customer technical lead who approves scope, rubric, and final interpretation.
Controls

How to keep the pilot credible

  • Freeze tasks, policies, target facts, findings, and scorer logic before K² runs.
  • Use the same generator and scorer in every arm.
  • Report context metrics separately from adversarial outcomes.
  • Keep commercial AI security tools out of public comparison unless they participate.
  • Report K² platform cost separately from downstream generator tokens.
Pilot arms and scoring

The pilot compares K²-supplied context against the customer's current context flow, using the same generator and scorer in every arm.

Evaluation arms

Current context flowHow the generator receives threat and target context today.
K² supplied contextThe same generator receives a cited plan from K² before generation.
Existing RAG/searchOptional control if the customer already has retrieval for threat context.

Scoring rubric

Targeting precision35%
Lineage rate25%
Reviewer time saved versus current flow20%
Regression-targeting hit rate10%
Policy-scope adherence10%
Typical pilot timeline

Keep the evaluation bounded enough to run and strict enough that the scorecard means something.

Week 1Freeze target, policies, past findings, candidate plan set, and scoring rubric.
Week 2Ingest threat, policy, target, and findings corpora into K² with metadata filters.
Week 3Run current context flow and K² supplied context against the same generator.
Week 4Review targeting, lineage, reviewer time, regression hits, cost, and residual gaps.
Recommended first target

Start with a target where context quality is already a review bottleneck. The support-copilot pattern is usually strongest because it combines retrieval, tools, modality, PII, and past mitigations.

Vertical options

Enterprise support copilotBest default: RAG, tools, screenshots, PII scope, and clear regression history.
Financial assistantUseful when severity definitions, synthetic PII, and policy-scope review are central.
SOC analyst agentUseful when tool agency, evidence lineage, and analyst-review handoff matter most.

Reviewer-time economics

25 plans/mo x 8 min saved$500/mo reviewer time value at $150 per loaded engineering hour.
50 plans/mo x 12 min saved$1,500/mo reviewer time value at $150 per loaded engineering hour.
100 plans/mo x 15 min saved$3,750/mo reviewer time value at $150 per loaded engineering hour.
After the pilot

If K² improves targeting, lineage, reviewer time, and regression targeting, the rollout scope is additional targets, corpora, teams, retention rules, and generator integrations. The customer owns all tasks, corpora, plan rows, and scorecards.

Success conditionHigher context fidelity and lower reviewer effort with the same generator.
Production scopeBounded rollout by target system, corpus, policy scope, and team.
Exit storyCustomer retains task definitions, scorecards, and configuration exports.
DisclosureNamed or anonymized case-study material is published only with approval.
Partner framing

If you build or sell an AI security product, K² operates below your generator. Your product remains the buyer-facing brand and K² supplies cited context beforehand.

Open Knowledge² website