Skip to main content

Collectors — overview & coverage

A collector captures AI activity from one layer of the environment and forwards it — with metadata and any local detection verdict — to the Lumen control plane. Collectors group by where the code runs:

GroupCollectorsRuns
Host-local (the hard part)Endpoint agent + Browser extensionOn the managed device
Instrument-the-codeApplication SDK, Agentic / MCP proxy, OpenTelemetryIn the built AI system
Network / platformGateways (Kong, LiteLLM, Portkey), Cloud (Bedrock + logs)At the API boundary / in the cloud account

Coverage matrix

SurfaceCollector(s)Inline enforce?
AI use in browsers (ChatGPT, Claude, Gemini, …)Browser extensionYes (redact/block in-page)
Internal applicationsApplication SDK + inline APIYes (pre/post model call)
Agentic / MCP traffic (client ↔ server)MCP proxyYes
Developer AI tooling (Claude Code, SDKs, CLIs)Endpoint agent proxy + hooksYes
API gateways (Kong, LiteLLM, Portkey)Gateway pluginYes
Cloud platforms (AWS Bedrock)Cloud ingestionLog/redact (async)
OpenTelemetry GenAI telemetryOTel receiverLog (observability)

Every collector can raise every detection class.

How an interaction flows

  1. Capture. A collector observes one AI interaction: the request (prompt, model, parameters) and/or the response, possibly streamed token by token.
  2. Enrich. Metadata is attached — user identity, device id, application or process, collector id, provider/model, tenant id.
  3. Evaluate. The collector's policy runs: access rules decide whether and how to process; prompt rules inspect the content. The local fast-path returns the inline verdict; deeper cloud detection runs asynchronously.
  4. Act. Log, Redact, or Block.
  5. Ship. The interaction and its verdict are spooled and drained to the Lumen cloud over outbound-only mTLS.
  6. Correlate. Findings power the console dashboards and export to the Wazuh Indexer.

Inline vs monitor

For Block/Redact the collector must sit in the data path and get a synchronous verdict (browser extension, SDK, MCP proxy, gateway, endpoint proxy). For Log-only it may observe out of band (cloud logs, mirrored traffic). The inline latency budget is a first-class design constraint: the verdict is local and takes single-digit milliseconds; the cloud never sits in the request path.