Docs/handbook/external standards map

External Standards Map

Handbook: Practitioner front door to assurance frameworks
Canonical crosswalks: content/standard/crosswalks/
Version: 1.0 | Updated: 2026-07-29


Purpose

Point engineering teams at the right external standard for the risk they are managing — without claiming certification or inventing a parallel dialect. AIES already holds machine-readable crosswalks; this page is the daily navigation layer.

Why

Teams either ignore external frameworks (“we have prompts”) or paste ISO/NIST logos on decks with no evidence. The correct posture: AIES artifacts as candidate evidence, assessors decide sufficiency.

Quick map (start here)

If you need… Open this Then this AIES surface
LLM app threats (injection, agency, DoS) OWASP LLM + Agentic crosswalk Security skill · Context pattern · MCP handbook
Org AI risk program NIST AI RMF crosswalk Portal / SoA · Migration bible · Evals gates
AI management system (cert path) ISO 42001 / 23894 / 42005 Enterprise pattern · Org playbook · Exceptions
Supply chain / SBOM / provenance SSDF / SLSA / OpenSSF DevOps skill · Release pattern · Hooks
MCP auth, A2A, GenAI telemetry Protocols & observability MCP when-and-how · LLMOps tracing
Legal high-risk classification (EU AI Act) Counsel + EU AI Act engineering map · Responsible AI Do not treat AIES as conformity

Evolving AI footprints (2025–2026) — how AIES absorbs them

Footprint What changed Where AIES covers it
Agentic OWASP Top 10 Goal hijack, tool misuse, memory poison, inter-agent trust OWASP agentic table in crosswalk · multi-agent threat-controls · agent definitions
MCP as default tool bus AuthZ, allowlists, confused deputy handbook/mcp-when-and-how · Tool Contract · empty mcpServers by default in kit
IDE agent skills (SKILL.md, AGENTS.md) Project skills are first-class Kit writes Cursor SKILL.md + root CLAUDE.md / AGENTS.md / Copilot instructions
Structured outputs / schemas Probabilistic text inside deterministic contracts Prompt templates · Tool contracts · Improper output handling (LLM05)
Eval-as-gate (not dashboards) Release blocking goldens / judges evals-as-quality-gates · EVAL family controls · continuous-eval
Prompt injection as data, not auth Untrusted retrieval cannot grant tools 09-enterprise-ai/security/prompt-injection-defense · Context pattern trust labels
Cost / unbounded consumption Token & tool budgets Containing token waste · LLM10 mapping · agent termination criteria

What is still thin (honest gaps)

Gap Mitigation today Do not
Full cookbook parity beyond kit stacks React/Next/Node/Python/Java/.NET/RN are kit-ready; domain cookbooks vary — see cookbook README status Pretend every cookbook is production-depth
EU AI Act conformity Informative engineering map — evidence only Claim CE / conformity via AIES
Live inter-org pilots (roadmap S1) Publish evidence notes when earned Market as “the industry standard” prematurely

Ease-of-use path (practitioners)

  1. /start → kit install → enhanceenhance-repo.prompt.md
  2. Pick a pattern (Planner / Context / Reviewer first)
  3. Run one lifecycle workflow
  4. Only then open portal/crosswalks for assurance

Anti-patterns

  • Citing “NIST compliant” because you read a blog
  • Enabling write MCP tools to “move faster” before OWASP LLM01/ASI02 review
  • Forking private prompt dialects instead of enhancing the kit against repo conventions

Checklist

  • Risk question mapped to one crosswalk row
  • Evidence artifact IDs listed (not just doc titles)
  • MCP still empty or allowlisted with data class
  • Eval gate exists for user-visible model behavior
  • Human gate named for irreversible effects

Changelog

  • 2026-07-29: Initial practitioner map — OWASP/NIST/ISO/SLSA/MCP footprints + honest gaps.