Artificial intelligence

Security awareness training MCP servers compared

Which security awareness training vendors publish an MCP server, and what each one can read, change and log, from each vendor's own documentation.

Short answer: As of November 2026, eight security awareness training vendors besides Kinds publish their own MCP server: Living Security, CanIPhish, Arsen, Phished, HailBytes, Fallax, Hook Security and Breach Secure Now. KnowBe4 and Adaptive Security have community-built servers but no first-party one we could find, and we found no MCP server for the training products of Hoxhunt, Proofpoint, EasyLlama and 12 other vendors.

We checked each vendor’s own documentation, announcements and code, plus the official MCP registry, on November 16, 2026. Every claim about another company comes from a primary source listed at the end.

What an MCP server is

The Model Context Protocol (MCP) is an open standard that lets an AI assistant use an outside product through a set of named tools. The vendor’s MCP server decides which tools exist. Some only read data. Others can change it. Our glossary explains what Model Context Protocol is in more detail.

The comparison at a glance

VendorBuilt byCan it change data?Sign-in
KindsKindsYou choose: Read, or Read, create and deleteOAuth: your usual Kinds login (Google, Microsoft or email). No API key
Living SecurityVendorYes, once an admin allows writesOAuth, or a client secret
CanIPhishVendorYes, including phishing campaignsOAuth, or an API key
ArsenVendorYes: launch campaigns, order testsSign-in, method not stated
PhishedVendorYes, including create and delete toolsSign-in, method not stated
HailBytes SATVendorYes: create and schedule campaignsAPI key
FallaxVendorTwo actions onlyOAuth: your Google or Microsoft account. No API key
Hook SecurityVendorNo, read-onlyAPI key
Breach Secure NowVendorNo, read-onlyExisting BSN credentials
KnowBe4Community projects onlyVaries by projectKnowBe4 API key
Adaptive SecurityCommunity project onlyNo, read-onlyAPI token
As of November 16, 2026, from each vendor’s own documentation or the project’s own code. “Not stated” means the source is silent.

Vendors with their own MCP server

Living Security

Living Security’s documentation describes a server that reads workforce risk, activity, training and phishing data. MCP access is off by default. Write tools stay blocked until an admin allows writes for a permission group, and actions that need approval in its Livvy assistant return “approval_required” over MCP and change nothing. Sign-in uses OAuth, with a client secret as the alternative. Its strongest point is the record: every tool call, including refused ones, goes into the organization’s audit log.

CanIPhish

CanIPhish’s help center describes a connector that reads the program and acts on it: creating campaigns, building employee lists, pulling reports and triaging reported emails. The assistant asks before any change, and a live phishing send needs a preview of the exact recipient list and a one-time confirmation. Sign-in is OAuth by default, or a pasted API key. It is limited to Professional and Enterprise plans, and to Account Admins and Super Admins.

Arsen

Arsen’s MCP page says the assistant can read your results and, with your approval, launch campaigns and order tests. One-off tests are paid in credits without a subscription, and the assistant shows the credit cost before every order. Arsen says email content and captured data are never returned. One control in account settings disconnects all AI access.

Phished

Phished’s connector is listed in Anthropic’s connector directory, made by Phished BV and marked “Anthropic verified.” The listing shows 96 tools, including create and delete tools, and says they give access to all endpoints in Phished’s API. It says sign-in is required but not how, and does not say whether changes are confirmed.

HailBytes SAT

HailBytes SAT is self-hosted, and each instance has a built-in MCP server with 14 tools. They list campaigns and results, and create campaigns, email templates, landing pages, target groups and sending profiles. It uses the same API key as the REST API, and a read-only user can list but not create. Every tool call is recorded in the audit log with the user, IP address and user agent.

Fallax

Fallax runs a narrow server: 12 tools, 10 of which only read. The two that act pause the simulation program or stop simulations for one person. Its documentation says nothing exposed through the server can create, launch or schedule a simulation. Sign-in is OAuth: you use the Google or Microsoft account you already use for Fallax, and there is no API key.

Hook Security

Hook Security announced its MCP connection on November 4, 2026. It is read-only: the assistant can read simulations, results and training but cannot change anything in Hook. An admin creates an API key, and an MSP key covers every client, including ones added later. Hook says it is included in all paid plans on its new platform.

Breach Secure Now

Breach Secure Now announced a read-only connector on November 10, 2026. Users sign in with their existing BSN credentials, and what they see depends on their role: MSP administrators see their managed clients, and employees see their own results. It names three assistants, including Microsoft Copilot.

Community-built servers

KnowBe4. We found no MCP server published by KnowBe4. Independent projects from Mirage Security and Servosity connect to KnowBe4’s Reporting API with an API key. Mirage Security’s lists a Platinum or Diamond subscription as a prerequisite. Servosity’s states it is not affiliated with, endorsed by or sponsored by KnowBe4.

Adaptive Security. We found no MCP server published by Adaptive Security. One independent project describes itself as beta, read-only and not affiliated with Adaptive Security. It uses an Adaptive API token.

Vendors where we found no MCP server

As of November 16, 2026, we found no public MCP server for the security awareness training products of Hoxhunt, Proofpoint, EasyLlama, Mimecast, SANS Security Awareness, Huntress, Arctic Wolf, NINJIO, usecure, Phin Security, CultureAI, Jericho Security, Riot, Infosec IQ or Cofense. For each, we searched the vendor name with “MCP” and “Model Context Protocol”, queried the official MCP registry and looked for vendor announcements. Not finding one does not mean none exists, and this changes quickly.

Huntress does publish an MCP server, but its announcement lists data such as agents, incident reports and invoices and does not mention awareness training.

What to ask before you connect any MCP server

  • Read or write? Can you limit a connection to reading, and who can grant more?
  • How do you sign in? A sign-in flow ties access to a person. A pasted key works until someone revokes it, so ask who holds it.
  • Is each change confirmed? Does the server show you exactly what will happen first, or does it rely on the assistant asking?
  • Is there an audit trail? Are calls recorded with the person, the assistant and what was done?
  • Where does the data go? Names, email addresses and results go to the assistant you connect, and so to the company behind it. Check your agreement with that company.
  • Who maintains it? A vendor server is supported by the vendor. A community server is not.

Where Kinds fits

The Kinds MCP server reads your learners, training, simulated phishing and compliance data. It can also take admin actions such as assigning training, sending reminders and removing learners. When you connect, you choose Read, or Read, create and delete. Sign-in is OAuth: you sign in the way you already sign in to Kinds, with Google, Microsoft or email, and approve the connection. There is no API key to create or paste. For each action, Kinds returns a preview of exactly what will happen, and you confirm before anything runs. Each action is written to an audit trail.

It is included in every account at no extra cost and stays off until an account admin switches it on. Any MCP client can connect, including an app you build. It has been tested with ChatGPT and Claude Code, and is designed for Claude, Cursor and VS Code as well.

Where others go further: the Kinds MCP server cannot launch phishing simulations or edit templates, and the CanIPhish, Arsen and HailBytes servers can start campaigns. It is not listed in Anthropic’s or OpenAI’s connector directories, and Phished’s is. And as with any of these servers, your data does not stay inside Kinds: it goes to the assistant you choose.

Related reading

Sources