Trust and security
A verifier you don't have to take on faith.
A tool that checks AI has to be checkable itself. These are the decisions we made so that it is.
Principles
The grader is not the one graded.
No AI in the decision path
Allow, block, write, learn: every decision is deterministic code over recorded data. An LLM may only suggest a query fix or help parse a badly formatted report, and anything it touched is labelled.
We don't investigate
Verdikor will never be an AI analyst. That is what lets it check any vendor's AI, including the platform's own, and stay on nobody's side.
Only people label outcomes
Whether an AI was right is judged only by human dispositions on real cases. Agreement between AIs never counts as correct.
Your knowledge stays yours
Lessons, sample queries and schema are plain YAML in your own git. Switch AI or SIEM and you keep all of it.
Evidence ledger
Change one row. Watch the chain break.
Each evidence item is canonicalised with RFC 8785 and hashed together with the hash of the item before it. A daily checkpoint is signed with Ed25519 and can be sent somewhere we don't control.
This demo computes real SHA-256 hashes in your browser. Edit an item and every hash after it stops matching.
Checkable without us
Open formats, closed code, independent checks.
Verdikor's code is not open source. What an auditor needs to check our output is.
Published specification
The evidence format, hash chain, checkpoints and verification report will be published under CC BY 4.0, so anyone can write their own checker.
Free verifier
A free, reproducibly built command-line verifier checks an export offline, without Verdikor running.
Third-party review
We plan an annual independent audit of the claim checker and the ledger, with a public summary, and code escrow for enterprise contracts.
Security by design
Built for data you can't let leave.
Runs where your data is
Self-hosted in your infrastructure, including networks without internet access and self-hosted models. You don't need to give us an LLM API key for the core to work.
Identity on every call
AI clients connect over MCP with OAuth 2.1. Every tool call carries a person, a tenant and an agent. There is no anonymous agent and no local stdio mode in production.
Tenant isolation in the database
Row-level security on every table, with a per-database option. Evidence, approvals and audit events are append-only for the application.
Least privilege toward your SIEM
Separate credentials per purpose: read, case read, case write, audit. Revoke the write key and every side effect stops.
These are design commitments for a product in development. We will publish what has been verified, and by whom, as it happens.
Report a vulnerability
Found something? Tell us first.
Email security@verdikor.com. Please don't access data that isn't yours, degrade the service, or use social engineering. Give us reasonable time to fix before you disclose.
Contact: mailto:security@verdikor.com Preferred-Languages: en, vi Canonical: https://verdikor.com/.well-known/security.txt
Early access
Run Verdikor on your own cases.
We're working with 3 to 5 design partners: MSSPs running several SIEMs, and banks or insurers using AI with any SIEM.