Greymatter.ioGreymatter.io
Start for Free
Greymatter.ioGreymatter.io Why Greymatter.io?
Pricing Start for Free
Learn
Resources/Blog

What is AI security

By Chris Holmes6 min read

AI security controls each model call and each tool call. The control runs in the data plane. The data plane is the path that already carries the call.

A firewall records an HTTPS connection. A cloud access security broker (CASB) records a session to a cloud application. An AI gateway at the edge records only the traffic that a team routes through it. None of these records names the person. None of them names a tool call inside the cluster. None of them names the token cost.

Programs adopt language models and agents faster than a review or a budget can follow. A program runs several frontier models and a sovereign model at the same time. Each model has its own account. Each application stores its own provider key. The application calls the model directly. The application writes only the logs that the team configured. Agents add tools through the Model Context Protocol (MCP). Each tool server is another endpoint. Each server authenticates the caller on its own.

You close this gap in the mesh you already run. You do not add a second product.

Gaps in the controls you already run

Five failures follow when each team calls a model on its own.

Each application stores its own provider key. The program cannot name the person who spent the key. The program cannot name the workload either.

A prompt can send personal data out of the enclave. A prompt can also send a health record, logistics data, or a classification marking.

A response can carry an injected instruction. A retrieved document can carry one too. The instruction can turn an agent against the person who asked.

An agent that can call a tool can delete data, write data, or run a command. Each MCP server is a new place that can act.

No record names the person who spent the tokens. A loop that does not stop can spend a month of provider budget in one day.

One question becomes many model calls and many tool calls. Most of those calls run after the person’s request returns. A control at the edge does not inspect those calls.

The control in the data plane

Greymatter governs that traffic in the data plane. Greymatter does not add an AI gateway. Greymatter does not require a software development kit (SDK). The application does not change.

The application keeps its OpenAI-compatible client. The application sends the call to an egress that the program declares. The call uses mutual TLS. The data plane carries the verified person with the call.

The egress reaches only the provider hosts that its routes name. The Console names each outbound call that skips the egress. The program can find a direct call to a model provider and close it.

Greymatter carries the verified person to the egress. Greymatter also carries that person to the tool server. The egress removes the person’s identity before the call leaves. The egress does not send the client’s token. An agent can keep working after the request ends. The record still names the person who asked.

Guardrails and data loss prevention (DLP) read the body once. They read it in both directions. Guardrails looks for attacks on the AI system. The modules map to the Open Worldwide Application Security Project (OWASP) Top 10 for large language model (LLM) applications. DLP finds personal data, health data, logistics data, and classification markings.

A team can set Guardrails to block one module at a time. Detect mode first records what enforce mode would stop. DLP records the finding. DLP does not block. A block on a valid patient record stops the work. A block on a shipping manifest stops the work. The record of that data must still exist.

Each person gets one daily token budget. The program can set a different budget for each group. Analysts and interns are examples. Routes that share one budget domain use one total for that person. The person can choose any model on those routes. Three separate budgets allow three times the tokens.

Every replica shares the count. More replicas do not increase the budget. The egress returns the quota error that an OpenAI-compatible client already handles. The egress reads the token count when the response ends. That call completes. The egress refuses the next call.

The same path covers tools. The MCP Router is a technology preview in Greymatter 2.5.0. The router offers many tool servers as one endpoint. Each role lists and calls only the tools that the role holds. The router does not show a tool that the role cannot use. The model cannot select a tool that the router does not send. The router sends the verified person to the tool server. The router does not forward the client’s token.

A record an assessor can test

Greymatter writes every model call to the audit record. Greymatter writes every tool call there too. The record includes the person, the findings, and the tokens. Greymatter stores a credential as a keyed fingerprint, or withholds the credential. A reader cannot replay a credential from the record.

Governance frameworks require that trail. The Organisation for Economic Co-operation and Development (OECD) AI Principles are one example. The National Institute of Standards and Technology (NIST) AI Risk Management Framework is another. ISO/IEC 42001 and the record-keeping duty in the EU AI Act are two more. Greymatter supplies the trail from live traffic.

The platform does not make a system compliant with those frameworks. The platform gives an assessor controls and records to test. The platform does not host the model. The platform does not score the quality of an answer. The platform does not replace the program’s review of a use case.

Guardrails and DLP detect known patterns. They reduce risk. They produce evidence. No inspection layer can promise that a model never misbehaves.

Ask Greymatter runs on this same path. The assistant calls its model through the platform egress. The egress charges the person’s token budget. The assistant’s tools call the Greymatter API as the person who asked. An assessor can watch these controls on a real assistant before the program approves its own applications.

Edge inspection and a CASB

An edge product measures the session that leaves the network. A CASB measures the session to a cloud application. Neither product measures an agent that calls a tool inside the mesh. Neither product measures a notebook that calls a model you run. Neither product measures a service that calls another service because a model sent it there.

Traffic often carries a shared service credential at a separate inspection point. The finding records the prompt. The finding does not name the person.

AI security is that governed path. The path carries a verified person, inspection in both directions, a budget, and a record with no secret. Greymatter puts that path in the data plane. The calls already go there.

See it running on your own infrastructure.

Reading is a start. A walkthrough with our team shows what Greymatter does in your environment.

Speak with a Greymatter.io Expert Start for Free How to Buy →

Additional resources

All Resources
Blog
By Chris Holmes
Blog · Nov 22, 2026AI security belongs in the data plane
Whitepaper
Download
Whitepaper · Nov 12, 2026One Path from People and Agents to Every Model and Tool
Platform
See AI Guardrails
Platform · AI GuardrailsGreymatter Guardrails and DLP run in the data plane beside every workload, with no extra hop. Keep agents inside their role, cap tokens per person, and record sensitive data by caller.