AI security fails when the inspection does not run on the call. Agents, notebooks, and services call models and tools from inside the enterprise. Most of that traffic never crosses a perimeter.
An edge service inspects the traffic that crosses the edge. A cloud access security broker (CASB) inspects the session to a cloud application. The caller at that point is often a shared service credential. An agent-to-tool call inside the cluster does not reach either product. A notebook call to a model you host does not reach either product. A service-to-service call does not reach either product.
Teams add an AI gateway, then a scanning API, then a separate gateway for each model provider. Each product adds a hop, a policy, and a log. Each product covers only the traffic that a team sends to it. Every application must change, or every route must point at the new product. A control that a team can forget to call is already off.
Greymatter runs the inspection in the data plane beside each workload. The data plane already decrypts the call. The data plane already holds the verified caller. Guardrails and data loss prevention (DLP) run there, on inbound calls, on outbound calls, or on both. They do not add a hop. You do not change the agent.
One pass and two products
Guardrails and DLP share one read of each message. They are not the same product.
Guardrails looks for attacks on the AI system. Its modules map to the Open Worldwide Application Security Project (OWASP) Top 10 for large language model (LLM) applications, 2025. The list includes prompt injection, supply-chain risk, and poisoning. It also includes unsafe output, excessive agency, and system-prompt leakage. It also includes cross-tenant retrieval, signals of a bad answer, and requests that exhaust tokens.
Guardrails starts in detect mode. Every rule runs. Guardrails blocks nothing. Each finding records whether enforce mode would block that call. A team turns enforcement on for one module at a time. The team sets the mode in the filter definition. The team does not wait for a new release. A team can enforce prompt injection and leave misinformation in detect mode.
DLP looks for sensitive data and records it. DLP has four modules. The modules cover personal data and secrets, health data, and logistics data. They also cover classification markings, such as a Controlled Access Program Coordination Office (CAPCO) banner. DLP reads requests and responses. DLP records a leak in a model answer the same way it records a leak in a prompt. DLP does not block. DLP does not redact. A block on a valid patient record or a shipping manifest stops the work.
An assessor uses the record later. The record shows which workload sent personal data to an outside model. The record shows which person the call served.
DLP reports a value only when the value is valid for its type. DLP does not report a card number that fails its check digit. DLP does not report a classification word in ordinary prose as a marking. Operators do not read a record that is full of false findings.
Greymatter runs inspection after it verifies the caller. Greymatter refuses an unauthenticated request before the scan. A caller header cannot turn the scan off. A caller header cannot change a score. A caller header cannot exempt the request. Greymatter records that attempt as a finding. The finding names the person Greymatter verified. The finding does not name a header that the application set.
Streams, sockets, and tool calls
A gateway that reads only one HTTP body does not inspect the other forms of AI traffic.
A model writes a streamed answer word by word. For each workload, a team chooses one of two modes. Buffer mode holds the answer until the check ends. Passthrough mode sends the answer as the model writes it. In enforce mode, passthrough mode stops the stream when Guardrails blocks.
Greymatter checks each WebSocket message on its own. Greymatter checks each Socket.IO message on its own. Chat applications and agents keep a connection open. The connection is not the unit of inspection. The message is the unit of inspection.
With Model Context Protocol (MCP) inspection on, Guardrails reads each tool call as a protocol message. Guardrails does not read the call as loose text. Guardrails reads the method, the tool, and every argument. Guardrails still finds a prompt injection that a caller hides in one argument.
Two parsers can read a repeated JSON key in different ways. Guardrails flags a repeated key and an ambiguous key. A hidden argument cannot pass through that key.
Guardrails returns a JSON-RPC error for a blocked MCP call. The error includes the rule number. The agent receives a protocol error. The agent does not receive a web page.
A team turns image inspection on as a separate choice. Guardrails then reads text inside an image. It uses the same injection rules and disclosure rules as other text. In enforce mode, Guardrails can block an image that hides a program.
Attackers change the spelling, the encoding, and the wrapper. Guardrails unpacks compressed content before the check. Guardrails removes invisible characters. Guardrails treats lookalike letters as the same letter. The check runs in the data plane. The content does not go to a scanning service.
The caller on each finding
An edge service often receives a shared service account. A CASB often receives one too. The product can report the text of the prompt. The product cannot name the person who sent it. The product cannot name the workload either.
Greymatter inspects the call after it verifies the caller. Each finding names the workload. The finding also names the person, when Greymatter has verified one.
A workload that runs for one person makes outbound calls with no inbound request. A personal analysis server is an example. A policy can name the annotation that holds the owner. Findings on those calls then name that person.
The inspection rule does not grant permission to use a tool. The MCP Router grants tools from the roles on the verified token. The router is a technology preview in Greymatter 2.5.0. The router applies those grants while Guardrails is in detect mode. The router still refuses a tool that the role cannot use. The router does not send client credentials to the tool server.
A team does not add inspection code to each service. One entry in the tenant onboarding policy selects workloads by label. The entry turns inspection on for inbound calls, outbound calls, or both. A new workload with that label gets the same inspection. The Console lists each listener where Guardrails is off. The Console lists each listener where DLP is off.
The record and the limits
Greymatter writes every Guardrails finding to the platform record. Greymatter writes every DLP finding to that same record. One event describes one call. The event names the caller and the place. The event records the decision, each rule, the scores, and the tokens.
Greymatter keeps only a short prefix of each body. A team can turn body capture off. Greymatter never stores a raw credential. No setting turns that protection off.
Greymatter keeps user names and token subjects. Those values identify a caller. A reader cannot replay them. Greymatter stores a bearer token or an API key as a keyed fingerprint. Greymatter withholds passwords, one-time codes, and private keys. It stores no fingerprint for those values.
A scan must not stop the data plane. When a scan passes its deadline, Greymatter forwards the traffic and records a coverage gap. A team sets the deadline for each product. Greymatter scans a large body up to the size that the team sets. Greymatter records that it did not scan the rest. In enforce mode, a team can block those bodies instead. On the workloads that must stop every unscanned body, a team can hold traffic until the scan decides.
The data plane refuses an invalid setting. The data plane keeps the last good configuration.
Guardrails reads text. Guardrails does not read pixels, audio, or video. Guardrails does not decide whether an answer is true. For misinformation, Guardrails flags signals that it can measure. Unsafe code and an invented package name are examples. A module looks for that OWASP risk in traffic. The module does not catch every form of the risk.
A governance program still needs a model registry, bias tests, and model cards. It also needs a decision about which use cases the program allows. Inspection of traffic cannot measure bias in a model. The data plane supplies the evidence those other systems do not collect. An edge service does not receive these calls. A CASB does not receive them either.

