Methodology
Discover. Govern. Test. Prove.
Every Korelex engagement runs the same four stages. Each one produces something you can check, and the last turns all of it into evidence a buyer, auditor or regulator can verify. You cannot test what you haven't found, and you cannot verify what you haven't tested.
01 — Discover
Find every AI component.
Every model, RAG pipeline, agent, tool graph, MCP server, third-party AI service and shadow deployment in the estate. Located, classified, written down — because you cannot test what you haven't found.
What we map
- Models and endpoints
- RAG pipelines and data sources
- Agents and tool graphs
- MCP servers
- Third-party AI services
- Shadow AI
What you get: the AI-BOM, your AI bill of materials. Everything after this is scoped against it.
02 — Govern
Write down what each control must do.
For each AI control — a guardrail, a filter, a permission boundary — we specify what it must refuse and what it must allow, precisely enough to test, and date it so drift shows. This is a specification we test against, not a policy we run for you.
What we specify
- Guardrail specification
- Refusal and allow cases per control
- Agent and tool permission scope
- Dated baseline
We verify the control. We never operate it.
What you get: a testable guardrail specification for every AI control in the AI-BOM.
03 — Test
Attack it like a real adversary.
Adaptive, multi-turn adversarial testing against the live system, through every legitimate channel. Confirmed exploits, not pattern matches. Then scheduled re-tests, so a fix that quietly regresses is caught.
What we attack
- Prompt injection
- Indirect injection
- Guardrail bypass
- Cross-tenant access
- Memory poisoning
- Secret exfiltration
- Scheduled re-test
- Regression proof on closed findings
What you get: root-caused findings, each carrying exactly one evidence label, and a scoped fix list.
04 — Prove
Keep the proof.
Every finding mapped to the frameworks your buyer and regulator already use, signed and dated, with drift checked against the AI-BOM. Each claim carries one label from our evidence standard, traced from the request to the regulator.
Mapped to
- OWASP LLM Top 10
- MITRE ATLAS
- EU AI Act
- GDPR
- RBI advisory
- India DPDPA
What you get: a signed evidence pack, security-review answers, a retest attestation and drift detection against the AI-BOM.
How it holds up
Four rules every finding has to pass.
A tester who also sells you the fix has a reason to find something. We sell neither, and every claim is held to these.
Re-run verbatim. Same status, same leak, or it isn't written down.
We record the refusal through every legitimate channel before claiming a bypass.
Read-only by default, hard request caps, PII redacted at the chokepoint.
Actual requests and transcripts ship with the report. Verify any claim yourself.
Put it to work
Run the four stages on your AI.
The AI Red Team runs all four before a deal or audit; KorelexOne runs them on every release.