Skip to content

Wiz MCP server: asking security questions in Claude and ChatGPT

6 min readFor cloud security leads and platform owners deciding how AI assistants should reach Wiz data

A Wiz MCP server lets Claude, ChatGPT or another MCP client ask about Wiz issues, vulnerability findings and configuration results inside a conversation. Settle three things first: that users sign in with their own Wiz account (or your SSO) so answers inherit their existing Wiz roles and projects, which scopes the tools need, and whether every tool only reads. Each answer should show the query it ran and what it counted as open or critical.

Why people look for a Wiz MCP server

Security teams already work in Wiz for triage. The questions that reach them from outside the team are different. A business unit leader asks how many critical issues are open in her cloud accounts. An engineering manager asks which of his team’s container images still carry a CVE with a fix. An auditor asks how the CIS pass rate moved since last quarter. Each is a short query against data Wiz already holds, and each usually turns into a screenshot or an export pasted into a chat thread.

An MCP server puts those questions where people already ask them. Claude, ChatGPT and Microsoft Copilot can call tools on a remote MCP server, so a person types the question and the assistant fetches the answer. The appeal is real, and so are the decisions it hides: whose credentials reach Wiz, which projects they cover, and how a reader knows the number matches what the console would show.

Access follows each user’s Wiz role

With user-level access, users sign in with their own Wiz account (or your SSO), and the MCP server inherits their existing Wiz roles and project scopes. A business unit leader sees the projects Wiz already gives her, and nothing more. Four settings are worth reviewing.

  • ScopesGrant read scopes only, such as read:issues and read:vulnerabilities, plus whatever the agreed questions need. An assistant answering questions has no use for a scope that changes issue status or resolves findings.
  • Project scopeWiz projects group cloud accounts and resources by team. If each business unit should see only its own cloud accounts, project-scoped roles keep that boundary inside Wiz, and answers follow it.
  • Who may askThe platform behind the MCP server decides which users may ask at all. Within that group, each user’s Wiz role decides what an answer can read.
  • RotationSign-in tokens need a named owner for the integration, an expiry policy and a tested revocation path, so a user removed from Wiz or from your SSO loses access on the next question.

Questions that work well

Good questions name a population Wiz already models: projects, severities, controls, frameworks and resources. They become reliable once the security team agrees a few definitions, because two correct queries can disagree on what open, critical or age means. The table below pairs common questions with the Wiz data behind them and the definition worth writing down first.

Notice that every row has a definition to agree. A count of open critical issues depends on which statuses count as open, whether issues on deleted resources are excluded, and whether critical means Wiz severity or a CVSS threshold. None of that is visible in the question, and an assistant that guesses silently will produce numbers that disagree with the console for reasons nobody can see.

QuestionWhat Wiz holdsDefinition to agree
How many critical issues are open per project, and how many are older than 30 days?Issues with severity, status, project and created dateWhich statuses count as open; age from the created date
Which projects have vulnerabilities with a fix available, detected more than 60 days ago?Vulnerability findings with fix availability and first detected dateWhether a reported fixed version counts as a fix
How has the CIS AWS benchmark pass rate changed month over month?Cloud configuration findings per rule and resourcePass rate by resource or by rule
Which controls raised the most new issues this month?Controls and the issues they raisedNew by created date, with or without reopened issues

What an answer should show

A count with no working shown invites the wrong kind of trust. Security numbers end up in risk registers and board decks, so an answer should carry enough for a reader to check it against the Wiz console without asking the person who ran it.

  • The readingWhich meaning of open or critical it used. When a word has several meanings, the answer should say which one it picked and offer the others.
  • The queryThe GraphQL query sent to Wiz, so an analyst can rerun it and compare.
  • Review stateWhether a named person on the security team has reviewed the logic behind this kind of answer, and when.
  • A useful declineWhen a user’s Wiz role does not cover a project, the answer should say so and say who to ask, instead of returning a smaller number as if it were complete.
Question inClaude orChatGPTCritical issuesopen per projectConnection rulesMay this user ask?Wiz GraphQL APIUser’s own access,read scopesAnswer with itsworkingReading, query,review stateQuestion in Claude or ChatGPTCritical issues open per projectConnection rulesMay this user ask?Wiz GraphQL APIUser’s own access, read scopesAnswer with its workingReading, query, review state
Wiz sees the requesting user, so their existing roles and projects decide the answer.

Joining Wiz with ServiceNow

Many teams raise ServiceNow tickets from Wiz issues. That makes some of the most useful questions cross-system: which critical issues have no open ticket, how long tickets for high-severity issues took to close by assignment group, and which issues are resolved in Wiz while the ticket stays open. Those questions need both systems in one answer, with each system’s query shown so the two sides can be checked separately.

The join needs an agreed key, usually the ticket reference that Wiz’s ServiceNow integration records on the issue. One issue can have several tickets, and one ticket can cover several issues, so counts need aggregating before they are compared. Both systems run with user-level access, so two users can see different findings and tickets for the same question. Say whose view it is on the answer.

Test it before you approve it

Pick three questions the security team already answers from the console, such as open critical issues per project, vulnerabilities with a fix older than 60 days, and the CIS pass rate for one account. Run them through the MCP server and through the console on the same day, and compare. Differences usually point at a definition, which is worth finding before a board deck does.

Then test the boundaries. Ask about a project outside the user’s Wiz role and confirm the answer declines instead of returning zero. Ask the assistant to resolve an issue and confirm nothing changes in Wiz. Remove a user’s access and confirm their next question is refused. Keep the results with the approval record, because the tools an MCP server exposes can change after you review them.

How Nexatron handles Wiz

Nexatron connects to Wiz through the Wiz GraphQL API with user-level access: users sign in with their own Wiz account (or your SSO), and Nexatron inherits their existing Wiz roles and project scopes. People ask in the Nexatron app or through Nexatron’s hosted MCP server from Claude, ChatGPT and Microsoft Copilot. Questions only read Wiz. Each answer shows the reading it used, with the other readings one click away, and the query it ran, and it reads Verified only once a named person on your team has reviewed the logic, with who and when shown.

A question Nexatron cannot answer under your rules is declined with the reason and what to do next. The Wiz connector page lists the data it reads, and the Wiz vulnerability management and compliance reporting pages give example questions and the definitions to agree first. The Wiz and ServiceNow page covers findings against tickets in one answer.

Claude is a trademark of Anthropic, PBC.

Bring three business questions

Explore your sources, definitions, and review criteria with our team. Please keep credentials and sensitive records out of the inquiry.