Skip to content

Standards and Incident Mapping

The mapping a security engineer needs to evaluate Boundary against the checklists already in use: OWASP Top 10 for Agentic Applications, the NSA's MCP security design considerations, the joint Five Eyes agentic-AI guidance, EU AI Act logging and human-oversight articles, SOC 2 evidence collection, and the public 2025 incident classes.

Every row maps to a delivered claim in the build-gated claims ledger, and every row carries the same ground rule: Boundary governs an action only when the route is forced through Boundary — the unrouted path is a bypass that deployment topology must remove. The mapping is architecture fit, not a compliance certification or guarantee.

The short version:

Framework hook Boundary answer Maturity
OWASP ASI01 (Agent Goal Hijack) Stage 1 trust check + adaptive termination; write-after-taint denial of the consequential routed action Trust delivered; taint profile preview, fixture-proof
OWASP ASI02 (Tool Misuse) Static policies + interceptors (including the Postgres AST guard) decide routed tool calls before execution Production on the routed MCP path
OWASP ASI05 (Unexpected Code Execution) Command Boundary decides routed commands before execution Delivered preview, routed paths only
NSA MCP CSI (mediating proxy, invocation logging) Production MCP JSON-RPC proxy + a structured, hash-verifiable decision record per governed verdict MCP production; records delivered
Five Eyes "accountability opacity" Hash-verifiable decision records, offline verify-record, explain, replay, evidence bundles Delivered
EU AI Act Art. 12 / Art. 14 Decision records; warn / escalate / require-approval verdicts Architecture fit for the 2027 obligations — not compliance
SOC 2 evidence requests boundary evidence bundle + boundary evidence verify (manifest-hashed, offline-verifiable) Delivered, local-only

The remaining seven OWASP ASI codes are listed as out of scope — mapped to nothing rather than stretched. Incident classes (production-database deletion, GitHub MCP write-after-taint exfiltration, MCP data exposure) are mapped to the pipeline stage that denies the routed shape, each with its own routed-only caveat; the worked example is the fixture-only boundary demo github-lethal-trifecta denial:

actual action: DENY
reason: lethal_trifecta_detected
upstream_called=false

Canonical mapping: docs/STANDARDS_MAPPING.md

Related references: