All skills
simota avatar

/canon

@e307415
by shingo imotasimota/agent-skills85 stars
15

Assessing standards, regulatory controls, and legal-document coverage with cited evidence and proposed wording. Use for OWASP/WCAG/SOC2/PCI/HIPAA or ToS/privacy/DPA reviews; not legal advice or code fixes.

Use this Skill: https://skilld.dev/gh/simota/agent-skills/canon

This session only. Nothing lands on disk.

referenceregulatory-vendor-risk-assessment.md

≈2.9k tokens on demand. Your agent reads this file only when SKILL.md points to it.

Third-Party Vendor Risk Assessment Reference

Purpose: Build a defensible third-party risk management (TPRM) program that keeps pace with SOC 2 CC9.2, PCI-DSS Req 12.8, HIPAA §164.308(b), ISO 27001 A.5.19-A.5.22, and GDPR Art. 28 expectations. Cover vendor inventory, tier classification, DPA / BAA signing, SIG / CAIQ questionnaire handling, SOC 2 report review, ongoing monitoring cadence, and subprocessor chain visibility. Every in-scope vendor has a tier, a contract, and a re-review date — no exceptions.

Scope Boundary

  • Canon vendor: vendor risk program design — inventory, tier policy, contract requirements, questionnaire process, SOC 2 report review, monitoring cadence, findings tracking.
  • Crypt (elsewhere): cryptographic architecture — asked when reviewing a vendor's encryption claims or key management design; Canon consumes the verdict.
  • Vigil (elsewhere): threat-hunt a vendor after a compromise; Canon runs the program that catches it earlier.
  • Sentinel (elsewhere): scan vendor SDKs you embed for vulnerabilities.
  • Cloak (elsewhere): privacy-specific vendor review (processor / sub-processor mapping under GDPR Art. 28) — pairs with vendor when personal data is in scope.

If the question is "should we sign with this vendor, what tier, and how often do we re-review?" stay in vendor. If it is "pen-test the vendor's API" or "review their encryption code", hand off.

Vendor Inventory

Every vendor that touches production systems, customer data, or regulated environments appears in a single registry. Shadow SaaS (expensed-as-reimbursement tools) is the biggest audit hole — sweep procurement and SSO logs quarterly.

Field Content
Vendor name Legal entity, not brand
Service description What they do for us, in one sentence
Owner Internal business owner + technical owner
Data shared Categories (customer PII, employee PII, CHD, ePHI, source code, telemetry)
Processing location Region(s) — feeds transfer-mechanism analysis
Tier Critical / High / Medium / Low (see matrix)
Contracts MSA, DPA, BAA, SCC, order form — with signed dates
Last assessment Date + outcome + evidence path
Next review Calendar date based on tier cadence
Subprocessors Their processor chain, last synced

Tier Classification Matrix

Tier Criteria (any one triggers) Controls required
Critical Processes regulated data (CHD/ePHI/special-category), part of the authentication path, outage would cause customer-visible incident in < 1h, or hosts a foundation model that powers a user-facing feature (treat as a subservice organization, see AI vendor section) SOC 2 Type II + pentest summary + SIG-Core, annual on-site or deep-dive, continuous monitoring, contractual right to audit
High Handles customer PII at rest, privileged infrastructure access (CI/CD, production IAM) SOC 2 Type II or ISO 27001, SIG-Lite, annual review, subprocessor sync
Medium Internal corporate data, limited PII, non-production access Attestation letter or CAIQ, biennial review
Low No data exchange, marketing tools, internal productivity Lightweight review at onboarding, event-driven re-review

Down-tiering a vendor requires written justification from the business owner and security lead. Up-tiering is automatic when the vendor's scope changes.

AI Vendor Risk (2026 must-have)

By 2026 the SOC 2 / TPRM community treats AI/LLM providers as a distinct vendor class. Two consequences for the inventory and tier matrix:

  1. An LLM provider that powers production behaviour is a subservice organisation, not a "tool". Their SOC 2 (or equivalent) becomes part of your own SOC 2 story; their controls flow into your CSOCs. Tier them at Critical by default.
  2. AI-specific questionnaire items are now standard alongside SIG / CAIQ — auditors expect to see answers on:
    • Model versioning policy (when did the version change; how is it pinned)
    • Prompt + completion logging retention and access controls
    • Training-data lineage and licence claims (Art. 53 EU AI Act for GPAI)
    • Hallucination handling and human-in-the-loop policy
    • Sub-processor chain for LLM hosting (cloud region, GPU tenancy, fine-tuning data residence)
    • AI literacy training records mirroring Art. 4 of the EU AI Act
    • Data-usage clause: is your input data excluded from provider training?

Customer procurement teams in 2026 increasingly refuse to open commercial conversations without an in-date SOC 2 Type II that covers the AI/LLM scope explicitly — the AI questionnaire is the new gating step before commercial alignment.

Contract Requirements

Contract When required Key clauses
MSA All vendors Liability cap, indemnity, termination, data return/deletion
DPA (GDPR Art. 28) Any processor of personal data Purpose limitation, subprocessor notice, audit right, SCC incorporation
BAA (HIPAA §164.314) Any business associate handling ePHI Permitted uses, safeguards, subcontractor BAA flow-down, breach notice within agreed window (24h under proposed 2026 Security Rule)
SCC 2021 Non-adequate country transfers Correct module (C2P / P2P / etc.), Annex I-III completed, TIA attached
Security addendum Critical / High tier Encryption requirements, incident notification window, pentest frequency, right-to-audit

Never onboard a vendor touching regulated data before the relevant contract is countersigned. "We'll get the BAA next week" is how HIPAA breaches compound.

Questionnaire Handling

Questionnaire Use when Pages Notes
SIG-Core (Shared Assessments) Critical / High tier, annual 800+ Deep; pair with evidence attachments
SIG-Lite High / Medium tier ~250 Good default; extensible
CAIQ (CSA) SaaS cloud providers ~260 Map to CCM controls; useful for cloud stack
Custom Never first choice — Reserve for gaps the standards miss

Store completed questionnaires in the evidence room with a response date, reviewer identity, and delta-vs-prior-year analysis. A questionnaire answered "yes" every year with no deltas is a signal the vendor is not tracking their own program.

SOC 2 Report Review Checklist

Do not merely file the PDF — review actively:

  1. Scope — does the report cover the system you consume? Trust Service Criteria included?
  2. Period — Type II period overlaps your audit period; gap-period letter if needed.
  3. CUEC / CSOC — Complementary User Entity Controls and Complementary Subservice Organization Controls: you are responsible for these. Enumerate them and map to your own controls.
  4. Exceptions — every finding by the vendor's auditor; assess residual risk.
  5. Subservice organizations — inclusive vs carve-out method; if carve-out, review the subservice's SOC 2 as well.
  6. Management response — serious exceptions must have remediation committed.
  7. Bridge letter — covers the gap between report end and your reliance date.

Output a short internal memo per report: scope match, period, CUECs owned, exceptions of concern, subservice posture, next review trigger.

Ongoing Monitoring Cadence

CRITICAL  →  quarterly control-health check (uptime + security posture)
          →  annual SOC 2 / pentest summary refresh
          →  real-time subprocessor notifications (DPA clause)
          →  immediate review on breach notice / CVE affecting vendor / M&A event

HIGH      →  semi-annual check-in
          →  annual SOC 2 / ISO cert refresh
          →  30-day window on new subprocessors

MEDIUM    →  annual attestation
          →  event-driven on change

LOW       →  biennial review
          →  event-driven only

Event triggers that force re-review regardless of tier: data breach at the vendor, ownership change (acquisition, bankruptcy), material change in the service, new regulatory requirement, repeated SLA misses.

Subprocessor Chain Visibility

GDPR Art. 28(4) and similar obligations require the prime processor to be liable for its subprocessors. Maintain a tree, not a list:

Your company
  └─ Vendor A (prime processor)
       ├─ Subprocessor A1 (hosting — AWS us-east-1)
       ├─ Subprocessor A2 (email — SendGrid)
       └─ Subprocessor A3 (analytics — Segment)
            └─ Sub-subprocessor A3.1 (warehousing — Snowflake)

Require vendors to publish and version their subprocessor list, and to notify you with an objection window (typically 30 days) before adding a new one. Capture sub-subprocessors for critical-tier vendors where the chain reaches regulated data.

Anti-Patterns

  • Relying on vendor marketing claims ("enterprise-grade security") without a SOC 2 or ISO 27001 report.
  • Signing a DPA without completing Annex III (list of subprocessors) — the agreement is incomplete.
  • Treating SOC 2 as binary pass/fail — ignoring exceptions, CUECs, and subservice carve-outs is the silent audit failure.
  • Using the same tier cadence for every vendor — critical vendors need quarterly eyes on them.
  • Shadow SaaS not in the registry — the vendor you do not know about is the one that leaks customer data.
  • Letting contract renewals roll forward without re-review — multi-year auto-renewal freezes vendor posture.
  • Accepting a Type I report where Type II is needed — Type I only covers control design, not operation.
  • Ignoring the bridge letter and relying on a SOC 2 whose period ended eight months ago.
  • Confusing "DPA signed" with "subprocessor chain mapped" — the second is ongoing work.

Handoff

To Cloak — subprocessor map under GDPR Art. 28, international transfer analysis for vendors in non-adequate countries, DPIA inputs for high-risk processing.

To Crypt — validate vendor cryptographic claims (encryption at rest, key custody, HSM attestation, BYOK / CMEK options).

To Sentinel — scan vendor SDKs / libraries embedded in your product for CVEs.

To Builder — vendor inventory registry implementation, webhook intake for subprocessor notifications, SSO-log to registry sweeper.

To Beacon — vendor health signal dashboards, SLA monitoring, alert paths on vendor status-page incidents.

To Scribe — vendor risk policy, tier classification standard, SOC 2 review memo template, subprocessor notice template.


Capability Detail (SKILL.md excerpt)

  • vendor_risk_assessment: Vendor inventory and tier classification, DPA/BAA/SCC contract gating, SIG/CAIQ questionnaire handling, SOC 2 report review (scope/period/CUECs/exceptions/subservice orgs), tier-driven monitoring cadence, subprocessor chain visibility, vendor_lock_in_scoring (data portability / API openness / migration cost estimate / contractual exit clause review), exit_playbook (per-vendor documented migration path with effort estimate + fallback alternatives + data-export contract terms), sla_machine_readable_ref (link to vendor-published SLA in parsed form when available; advisory-only when vendor publishes only PDF), deprecation_calendar_ref (link to vendor deprecation announcement feed when available). v8 fold-in: lock-in/exit fields are advisory only — never block adoption since most vendors do not publish machine-readable SLA / deprecation calendars (omen v8 FM-V8-12 RPN 280 vendor prerequisite tyranny).

Source: SKILL.md on GitHub

1 warning13d5 checks · Risk SAFE
  • Gen Agent Trust Hub13d

    The 'canon' skill is a comprehensive framework for assessing software against security, accessibility, quality, and regulatory standards. It provides extensive reference documentation and templates for auditing projects. The skill is verified as safe, with all identified behaviors—such as the use of well-known auditing tools and policy-as-code execution—being standard practices within its functional domain.

  • Socket13d

    No alerts

  • Snyk13d

    Risk: LOW · No issues

  • Runlayer6mo

    2/6 files flagged

  • ZeroLeaks5mo

    Score: 93/100 · 2 sections analyzed

Signed by skilld at e307415. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub 2 days ago.

Activeupdated 2 weeks ago

README badge

README badge for simota/agent-skills/canon