Independent verification for AI in the SOC
Your SOC's AI writes convincing reports. Verdikor shows which parts are true.
Verdikor sits between any AI and any SIEM. It checks each query before it runs and traces every claim back to raw evidence after. No LLM in the decision path, so the checker never shares the blind spots of what it checks.
- No LLM in the decision path
- Runs in your infrastructure
- Any SIEM, through plugins
-
backup.exe on BKP-01 read 37 files in /srv/finance between 01:00 and 01:40.
-
The host connected to 185.220.101.7, a Tor exit node.
-
The platform's AI Summary says this resembles ransomware staging.
-
The hash of backup.exe matches its CMDB record. Owner: Infra.
Recommendation: isolate BKP-01. Withheld. Its basis did not hold.
Sent to analyst for review Pending
Built for the stack you already run
SIEM Microsoft Sentinel · Splunk · Elastic · Google SecOps · Cortex XSIAM · and more
AI Any MCP-compatible AI assistant or agent
The gap
Every SOC has bought AI. Almost none will let it act alone.
An AI report reads well. The analyst still can't tell which query ran, whether that IP is really in the data, or which line is one AI quoting another. So they trust it blindly or investigate again from scratch.
Source: State of AI in the SOC 2026, Prophet Security and ViB, August 2026. Survey of 250 security leaders and analysts.
How it works
Three checks. All deterministic.
Before the query runs
Parsed in the SIEM's own grammar, checked against the tenant's real schema, scored for cost and risk.
Query check 02After it runs
Every result becomes hashed evidence. Every claim must cite it. Entities are looked up in the raw rows.
Evidence and claims 03What the AI remembers
The AI only proposes lessons. People approve them by risk tier, with a backtest, an expiry and a way back.
Reviewed knowledgeFor the whole team
New analysts learn your SOC, not a generic one.
Every SOC runs differently: its own assets, exceptions, naming and past decisions. The knowledge your team approves for the AI is the same knowledge a new analyst can ask about, with every answer tied to its source.
- Approved lessons, sample queries and your real schema in one place
- How your team decided similar cases before
- Plain answers when something isn't covered yet
Expected. The nightly backup job reads /srv/finance between 01:00 and 02:00 on BKP-* hosts. Outside that window, or on any other host, it is not covered.
- Lesson L-112 approved by SOC lead · expires in 54 days
- 12 past cases closed as benign by analysts · all agree
- Asset owner Infra · from CMDB
Every answer points to an approved lesson, a past analyst decision or your measured schema. If nothing approved covers the question, Verdikor says so.
Expected. The nightly backup job reads /srv/finance between 01:00 and 02:00 on BKP-* hosts. Outside that window, or on any other host, it is not covered.
- Lesson L-112 approved by SOC lead · expires in 54 days
- 12 past cases closed as benign by analysts · all agree
- Asset owner Infra · from CMDB
In your network logs the remote address is remote_ip. dst_ip does not exist in this tenant, and it is the most common mistake in new queries.
- Schema measured 6 hours ago
- Sample query Q-031 reviewed by the team
- Common mistake blocked 9 times this month
Alerts from the scanner subnet during the Tuesday scan window are closed as benign, with the scan ticket linked. Alerts outside the window are investigated as normal.
- Lesson L-064 approved by 2 people · expires in 30 days
- 41 past cases closed as benign · 95% agree
- Runbook vulnerability scanning
No approved knowledge covers this. In your policy, excluding a host from a detection needs two approvers. Propose it and your SOC lead will review it.
- 0 approved lessons for this host and detection
- Policy exclusions need 2 approvers
Illustrative example
Try it
Pick your SIEM. Watch a bad query get caught.
The AI wrote these four queries for the same question. Verdikor checks each one in that SIEM's own grammar, against what the tenant really has.
More on query check1 DeviceNetworkEvents 2 | where Timestamp > ago(90d) 3 | where RemoteIp == "185.220.101.7"
Blocked UNKNOWN_FIELD line 3
Column names are case-sensitive. RemoteIp does not exist; the column is RemoteIP.
Warning line 2: lookback is longer than this table's retention.
1 index=* sourcetype=pan:traffic 2 dest_ip=185.220.101.7 earliest=-30d 3 | stats count by src_ip, dest_port
Needs approval COST_HIGH line 1
index=* searches every index for 30 days. That is above the share of quota policy allows one query to use.
Suggested: index=netfw, where this sourcetype was seen in the last 7 days.
1 FROM logs-network-* 2 | WHERE destination.ip == "185.220.101.7" 3 | STATS hits = COUNT(*) BY host.name
Allowed fields exist · cost within budget · risk low
Not every query is a problem. Most pass straight through, and the run is recorded as evidence.
1 dataset = xdr_data 2 | filter dst_ip = "185.220.101.7" 3 | fields agent_hostname, dst_ip
Blocked UNKNOWN_FIELD line 2
dst_ip is not a field of xdr_data in this tenant. Schema measured 6 h ago.
Suggested: action_remote_ip
Illustrative examples
Where it sits
One core in the middle. Nothing changes at either end.
Any AI
- AI SOC analysts
- Coding and chat assistants
- Platform agents
- Your own agents
Verdikor
- Query check
- Evidence ledger
- Claim check
- Reviewed knowledge
Any SIEM
- Microsoft Sentinel
- Splunk
- Elastic
- Google SecOps
- Cortex XSIAM
People
Approval inbox · weekly review · report for auditors
Everything vendor-specific lives in a plugin: grammar, schema, how to run a query, how to write to a case. Adding a SIEM means writing a plugin, not changing the core.
Who it's for
Teams that have to answer for what their AI did.
MSSPs and MDRs
Several SIEMs, many customers, more than one AI. One verification layer across all of them, and a report you can hand to your customer.
Regulated industries
When an auditor asks what the AI did, based on what, and who allowed it, answer with a signed report whose hash chain checks itself.
SOC leads growing a team
New analysts get up to speed on how your SOC works by asking the knowledge your team has already approved.
SOC teams building their own
Query lint against your tenant's real schema and a tamper-evident ledger of every AI tool call.
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.