Continue normally
The DLP service allows the content. AI Warden still applies native AI policies, budgets, model routing, MCP controls and request logging.
Platform · ICAP / DLP
Do not replace your DLP strategy. Put it inside the AI path.
AI Warden can call ICAP-supported DLP platforms, including common enterprise DLP and SSE inspection services where supported, before AI requests leave and before responses return. Your DLP outcomes can allow, modify, block or trigger session-level controls according to policy.
Request and response enforcement
Inspect before egress. Inspect before delivery.
AI risk is bidirectional. Users can send sensitive data to a model or tool, and models or tools can return content that should not be delivered. AI Warden gives ICAP inspection points on both sides.
LLMPromptsInspect prompt text, attachments and request metadata before provider egress.
LLMResponsesInspect model outputs before the client, agent or application receives them.
MCPTool callsInspect JSON-RPC arguments and tool requests before upstream tool execution.
MCPTool resultsInspect tool responses, files and generated content before returning to the agent.
Inspection lifecycle
ICAP decisions become AI Warden enforcement events.
AI Warden can sit between employees, apps, hosted agents, LLM providers and MCP servers. The gateway resolves identity and policy, sends the relevant body and context to ICAP where configured, applies the DLP decision, then continues with AI Warden scanners, routing and audit.
- Allow: continue through AI Warden policy, budget, routing and response inspection.
- Modify: forward the redacted or rewritten body when the ICAP service supplies one.
- Block: short-circuit the AI request with a clean policy response and audit outcome.
- Error or timeout: fail closed where configured so unavailable inspection does not become silent bypass.
1. receive user prompt / upload / MCP tool call 2. resolve user · agent · provider · MCP server · policy 3. inspect request allow | modify | block | error 4. enforce block, forward modified body, or continue 5. upstream LLM provider or MCP server 6. inspect response review generated response or tool result 7. deliver allowed/modified content or policy denial 8. audit request log with inspection outcomes
AI-aware actions
DLP decisions can control the AI session, not just a single packet.
A hosted chatbot receiving a file upload can send that content to ICAP. If the DLP response is modify or block, AI Warden can decide what happens next in the AI experience.
Warn or restrict
Deliver the modified body, warn the user, remove sensitive content, or disable selected tools such as internet search, external connectors or file export for that session.
Stop or contain
End the chat session, block the tool call, quarantine an agent, require approval, notify SOC, or auto-disable a risky user or agent after repeated events.
Fail-safe default
If ICAP does not respond, AI Warden can fail closed.
For regulated enterprises, an unavailable inspection path should not silently become no inspection. AI Warden can treat ICAP timeouts, connection failures, malformed responses and configured body-size violations as deny outcomes, with audit detail for the failed inspection attempt.
Behaviour-based controls
Repeated ICAP outcomes become a user, agent and workflow risk signal.
AI Warden knows the authenticated user, delegated user, hosted agent, MCP server, provider, model and session behind each event. That makes it possible to act on patterns, not just individual blocks.
- Alert SOC when a user or agent receives a high number of ICAP block or modify decisions.
- Trigger review or disablement for an agent that repeatedly attempts to disclose customer-identifiable data.
- Move a session into restricted mode after the first serious DLP event.
- Disable only the risky tools, such as internet search, external MCP connectors or file export.
customer-insights-agent12 blocks
external-search tooldisabled
finance-expense-agent0 blocks
Example action: disable internet search for the active session, notify SOC and route the agent owner through review.
Audit evidence
Governed AI requests can show whether ICAP inspected them and what happened.
Request logs can retain the gateway outcome, AI Warden policy outcome, ICAP request decision, ICAP response decision, service status, latency, flags, principal, delegated user, agent and target system.
request-log outcomeDenied by ICAP
traffic_type: mcp principal: aiwadmin agent: aiw-demo-customer-insights-assistant tool: customer_profile_export icap.request: allow · 204 · 24ms icap.response: blocked · 200 · 31ms action: disable external export tools for session outcome: denied before client delivery evidence: queryable · exportable
Use your DLP investment
Run AI Warden policies and enterprise DLP together.
AI Warden does not ask regulated customers to trust a separate AI-only policy layer instead of their DLP programme. It lets both operate in the request path, with configurable fail-closed enforcement and AI-aware session actions.