Skip to content

Attack-control mapping

ARCHITECTURE.md enumerates the enforcement points; this page relates selected Microsoft recommendations to those mechanisms.

The primary scenario is Microsoft’s AI memory and context poisoning technique. The same failure can begin as indirect prompt injection, become persistent data poisoning, and exploit weak retrieval-store integrity. The linked Microsoft catalog maps those stages to OWASP LLM01, LLM04, and LLM08 and to MITRE ATLAS context and RAG poisoning techniques.

The normative mapping is in SECURITY.md.

This is a partial mapping, not full implementation of Microsoft’s controls. Ziggurat uses strict record schemas but retains free-text bodies, preserves hostile instructions as evidence, and does not provide automatic purges, semantic sanitization, or drift monitoring.

Memory-poisoning controlEnforcement point or workflow
Source approvalInbox boundary, Bronze store, Gold eligibility
ProvenanceEvidence validator, index verifier
Memory write governanceReceipt verifier, external signing workflow
Strict record and proposal schemasProposal contract, corpus collector, index verifier
Review and diff transparencyHuman review packets
Presentation sanitizationHuman review packet rendering
IntegrityPage canonicalizer, receipt verifier, index verifier
IsolationProfile builders
RevalidationGold eligibility, index verifier
Versioning and rollbackOperator-managed Git history and index rebuilds
Least privilegeProposal store, refine reference builder, MCP server
Suspicious instruction handlingRefine reference builder, index verifier
Resource boundsAdapter transport, refine reference builder, MCP server

Gold admission fails closed unless every check in the Gold eligibility checklist passes.

Proposal corruption and index mismatches fail closed as specified in SECURITY.md.

These controls do not establish semantic support, factual truth, humanity, review quality, or process or tenant isolation; see the residual risks.