AI security

Decide what reaches a model before it leaves your network

Classic DLP was built for files leaving the network. A prompt is an authorised request to an authorised service, with the data in the body. MERIDVAR inspects that request on your network and can refuse it.

  • Inspected on your network
  • No provider call on refusal
  • Findings stored masked
  • Rules-based and deterministic
Gateway policyIllustrative

Policy

Edited in the console. Enforced on every routed request.

Shadow AI discoveryOffObserveEnforce
National identification numbersBlockRedactAllow and log
Payment cardsBlockRedactAllow and log
Email addressesBlockRedactAllow and log
Residency when the region is unknownFails closed
Team budget capHardSoft

Admission control

A gateway that is allowed to say no

Seventeen stages run on every routed request. If one of them refuses, the request does not reach the provider and nothing is sent. The refusal is recorded with its reason.

How MERIDVAR controls an AI requestRequests from people, applications, agents and MCP servers pass through the MERIDVAR admission gateway on the customer network. An admitted request reaches the model provider. A refused request stops at the gateway and the provider is never contacted. Every decision is written to a hash-chained evidence log.Your networkModel providersPeopleApplicationsAI agentsMCP serversMERIDVAR admission gatewayA pipeline of seventeen stagesidentitypassshadow-aipassprompt-dlprefusedthreatskippedresidencyskippedvendor-riskskippedbudgetskippedAnd ten more stagesAnthropicOpenAIAzure OpenAIxAIAdmitted: one callRefused: no callGovernance evidence logClassified9f2c…a71eRevieweda71e…03bdApproved03bd…c45aRefusedc45a…e210
Illustration of documented behaviour. A refused request never reaches the model provider, and every decision is written to an append-only, hash-chained log that names who made it, when and on what basis.

Refusal means zero calls

Tested against a provider that logs every request it receives.

The whole request is read

Prose, content blocks, system messages, tool call arguments and tool results.

Findings are masked

A raw identifier is never stored.

Controls

Policy you edit in the console, enforced by the same records

Shadow AI discovery
A list of sanctioned providers and a mode of off, observe or enforce. Upload proxy or DNS logs to see services that never touched the gateway.
Prompt DLP
Ten categories of structured identifiers, including national identification numbers, payment cards, email addresses, phone numbers, keys and tokens. Each category is set to block, redact or allow with logging.
Threat detection
A threat detection policy applied by severity at the gateway.
Data residency
Declared regions and a policy for each team. It fails closed when the region is unknown.
Vendor risk requirements
Data processing agreements, the age of the assessment and training on customer data, enforced at the call.
Identity at the gate
A bearer token from your identity provider can be required for admission. Each monitored person has a token that can be rotated and revoked.
Agent identity
Agents sign requests with Ed25519 keys, so a machine caller proves possession of a registered key and every agent request is attributable to a known identity.
MCP inventory
MCP servers and their tools are declared and risk-rated. A tool that can delete or run commands raises the risk of every AI system linked to it.
Endpoint agents
Endpoint agents, with installers for Linux, macOS and Windows, report the local AI tools they find. Presence of an agent can be required for admission.
Agents and MCP serversIllustrative

MCP server inventory

Declared servers and the risk of the tools they expose.

ToolCanTool risk
tickets.searchReadLow
tickets.updateWriteModerate
tickets.deleteDeleteHigh

A tool that can delete or run commands raises the risk of every AI system linked to it.

Security records

Every refusal leaves a trail you can check

  • Findings from every scan, with masked samples.
  • An audit trail in which each entry carries a hash chained to the one before, with a button to verify that it has not been altered.
  • Log shipping to a webhook you control, and Prometheus metrics.
  • Sign-in lockout after repeated failed attempts, for each account and each address.

The product views on this page are illustrative, with invented example data.

How it fits

Precise by design, and built to sit beside the security stack you already run

Deterministic detection

Prompt DLP is rules-based and matches structured identifiers. The same request always gives the same decision, and every refusal names the rule behind it.

Enforcement where traffic is routed

Controls apply to requests that pass through the gateway. Uploaded logs, endpoint agents and declarations extend visibility to the rest.

A complement to EDR and SSE

MERIDVAR adds AI-specific admission control and governance at the point of the request, alongside your endpoint and network security.

Watch a request being refused on your own network

A working session with the founder, on a host you control.

What happens next

  1. You send the formYour request goes straight to the founder.
  2. We reply by emailWe agree a time and what you want to see.
  3. We run a working sessionWe install the appliance with you, route a test request through it and read the evidence it leaves.