Skip to main content

OWASP agent, MCP, and RAG controls

All Dockprom-AI agents, skills, MCP servers, RAG pipelines, and execution containers must follow these references:

The MITRE webpage returned HTTP 404 during repository validation on 2026-10-30; the official ATLAS dataset remains the fallback source for AML.M0024.

Minimum controls​

ComponentMinimum controls
AgentBounded immutable goal, unique identity, least privilege, budgets, goal-drift detection, kill switch.
SkillVersioned manifest, strict schemas, declared tools/resources/egress, no embedded secrets, SBOM and digest.
MCP serverAuthenticated transport, origin/session binding, tool allowlist, parameter validation, authorization per call, complete audit.
RAGTrusted ingestion, provenance, tenant isolation, poisoning/injection scanning, relevance threshold, citations, revocation.
LoggingStructured sanitized events, trace correlation, time sync, access control, retention, no secret or prompt leakage.
ContainerNon-root, least capabilities, read-only filesystem, resource limits, restricted egress, no unrestricted Docker socket.
Model operationsPinned model/prompt/policy versions, evaluation gates, telemetry, canary deployment, rollback.

Required security telemetry​

Every agent run must correlate:

agent.run
-> model.operation
-> rag.retrieve
-> policy.decision
-> mcp.tool.call
-> human.approval
-> action.execute
-> action.verify
-> rollback or completion

Record decision summaries and evidence references, not hidden chain-of-thought. Redact credentials, PII, prompt/response content, tool payloads, and sensitive retrieved text before telemetry leaves the process.

The complete control matrix and release checklist are maintained in docs/SECURITY_CONTROL_REFERENCES.md.