All skills
simota avatar

/cloak

@35ffd55
by shingo imotasimota/agent-skills85 stars
15

Engineering privacy and data governance: PII detection, data flow mapping, consent patterns, GDPR/CCPA-compliant implementation, DPIA. Use when privacy-by-design is needed.

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

This session only. Nothing lands on disk.

referenceccpa-cpra.md

≈2.6k tokens on demand. Your agent reads this file only when SKILL.md points to it.

CCPA / CPRA Compliance Reference

Purpose: Implementation guidance for the California Consumer Privacy Act (CCPA, 2018) as amended by the California Privacy Rights Act (CPRA, 2020, fully effective Jan 2023) and the 2026 ADMT/risk-assessment regulations. Covers consumer rights automation, opt-out signal handling (GPC), Sensitive Personal Information (SPI) limitation, and the service-provider/contractor/third-party distinction that determines contractual obligations and disclosure thresholds.

Scope Boundary

  • cloak ccpa: California-specific implementation (consumer rights endpoints, GPC honoring, "Do Not Sell or Share" link, SPI limit-use mechanism, service-provider contracts).
  • cloak gdpr (sibling): EU regime — opt-in/lawful-basis model. CCPA is opt-out by default; do not conflate.
  • cloak pii (sibling): detection. CCPA ccpa consumes a PII inventory but enforces rights/disclosure on top of it.
  • cloak consent (sibling): consent-capture UI patterns. CCPA opt-out flows live here when scoped to CA users.
  • cloak dpia (sibling): CCPA risk assessments (selling/sharing PI, SPI, ADMT, biometric) are formally documented in dpia; ccpa only flags when one is required.
  • canon gdpr / standard compliance (elsewhere): standard checklists across regulations. ccpa produces actionable code, not audit narrative.
  • crypt (elsewhere): encryption primitives backing CCPA's "reasonable security" duty. Cloak references; Crypt designs.
  • canon (elsewhere): legal review of Privacy Policy clauses, MSA addenda, DPA language. Cloak emits the technical surface; Canon approves the prose.

Workflow

DISCOVER  →  identify CA-resident data subjects, SPI fields, sale/share flows,
          →  third-party tags, ADMT touchpoints

CLASSIFY  →  bucket processors as service provider | contractor | third party
          →  flag SPI categories triggering limit-use right

MAP       →  trace PI → service providers (contract-bound) vs third parties (sale/share)
          →  GPC ingress → consumer-rights state store → downstream propagation

ASSESS    →  rights coverage (know/delete/correct/opt-out/limit), GPC honoring,
          →  visible confirmation, 2026 risk-assessment triggers, ADMT readiness

REMEDIATE →  /ccpa/opt-out, /ccpa/access, /ccpa/delete, /ccpa/correct, /ccpa/limit-spi
          →  GPC header parser + visible "Opt-Out Request Honored" UI
          →  contract templates (1798.140(ag) service provider, (j) contractor)

VERIFY    →  45-day SLA timers, authorized-agent flow, minor opt-in (<16),
          →  do-not-sell link present site-wide, financial incentive disclosure

Consumer Rights (CCPA + CPRA)

Right Statute Trigger SLA Notes
Know (categories + specific pieces) §1798.100, §1798.110 Verifiable consumer request 45 days (+45 ext.) 12-month lookback waived by CPRA — now full history
Delete §1798.105 Verifiable consumer request 45 days 9 enumerated exceptions (security, legal, free speech, etc.)
Correct §1798.106 CPRA addition 45 days Must propagate to service providers
Opt-out of Sale or Sharing §1798.120, §1798.135 Single click; GPC signal Effective immediately; 15 days to propagate "Sharing" = cross-context behavioral advertising
Limit Use of Sensitive PI §1798.121 SPI processed beyond §7027 narrow purposes Effective immediately Required link if SPI used for inferences
Non-discrimination §1798.125 Any rights exercise — Financial incentives allowed if disclosed and reasonably related to value
Opt-out of ADMT 2026 Regs §7220 Significant decisions, profiling Pre-use notice + opt-out + appeal Effective Jan 1 2026
Data portability §1798.100(d) Subset of right-to-know 45 days Machine-readable, structured

Sensitive Personal Information (SPI)

Category Examples Limit-Use Trigger
Government IDs SSN, driver's license, passport, state ID Always
Account credentials Username + password, security Q&A Always
Financial Account number + access code, debit/credit + CVV Always
Precise geolocation <1,850 ft (~564 m) radius Always
Racial / ethnic origin Self-reported or inferred Always
Religious / philosophical beliefs — Always
Union membership — Always
Communications content Mail, email, text content (non-addressee) Always
Genetic / biometric / health DNA, fingerprint, face geometry, medical Always
Sex life / sexual orientation — Always

Narrow-purpose exemptions (§7027(m)): perform requested service, prevent fraud/security incidents, ensure physical safety, short-term transient use, perform internal services, comply with law, verify/maintain quality. Anything else triggers the Limit-Use link.

Service Provider vs Contractor vs Third Party

Entity Definition Key contract terms Disclosure obligation
Service Provider §1798.140(ag) Processes PI on business's behalf for business purpose No sale/share, no combining across clients, certify compliance, allow audits Disclosed but not "sale"
Contractor §1798.140(j) Same role, distinct definition for non-processor recipients Same as service provider Disclosed but not "sale"
Third Party §1798.140(ai) Anyone receiving PI not as service provider/contractor Sale/share — opt-out applies Disclosed as sale/share

Failure to bind = automatic third-party classification. Sephora ($1.2M, 2022) was the first major settlement: pixel/cookie sharing without opt-out treated as sale.

GPC (Global Privacy Control)

Aspect Requirement
Detection Read Sec-GPC: 1 HTTP request header on every page load
Honor scope Treat as opt-out of sale/sharing AND opt-out of cross-context behavioral advertising
Propagation Apply to known-user identity if logged in; persist across sessions
Visible confirmation (2026 CA Regs) Display "Opt-Out Request Honored" or equivalent when GPC processed
Conflict with cookie banner GPC overrides ambiguous banner state; never re-prompt
Browser support (2027) AB 566 — all CA-distributed browsers must include GPC by Jan 1 2027

Twelve states require GPC honoring by Jan 1 2026 (CA, CO, CT, DE, NJ, NH, OR, MN, MT, TX implicit, others). Cloak's ccpa recipe assumes CA semantics by default; widen scope when other states apply.

2026 Regulations — ADMT, Risk Assessment, Cybersecurity Audit

Obligation Trigger Deliverable
ADMT pre-use notice Significant decision use Plain-language description, logic, opt-out link
ADMT opt-out Same Functional channel — appeal to human review
ADMT access Same Decision-specific output, key reasons, role of input data
Risk Assessment Selling/sharing PI, SPI processing, ADMT, biometric processing, training AI on PI Documented benefits/risks; submit abbreviated version to CPPA
Cybersecurity Audit Threshold business + high-risk processing Annual independent audit; certified to CPPA
DROP — DELETE Request and Opt-out Platform Data brokers Single-channel deletion via centralized platform

Enforcement Snapshot

Penalty Amount
Per unintentional violation $2,663
Per intentional or minor-related violation $7,988
Statutory damages (data breach, private right) $107–$799 per consumer per incident

Notable actions: Sephora $1.2M (2022, sale without opt-out), DoorDash $375K (2024, sale to ad cooperative), Tractor Supply $1.35M (2024, GPC failure), Honda $632K (2024, dark patterns + opt-out friction).

Anti-Patterns

  • Treating "Do Not Sell" as a sale-only flag — CPRA added "or Share" for cross-context behavioral ads; pixel/tag-based ad personalization is sharing even if no money changes hands.
  • One-link policy footers — both "Do Not Sell or Share" and "Limit the Use of My Sensitive PI" must be present (or one combined link with both rights). Hiding either is a violation.
  • Asking the user to log in to opt out — opt-out must be available without authentication; only verification of identity-tied requests (delete/access/correct) requires login.
  • Honoring GPC silently — 2026 CA regs require visible confirmation. Set data-gpc="honored" and surface a banner.
  • Skipping service-provider contracts — without the §1798.140(ag) contractual flow-down, the vendor is a third party and every transfer is a sale.
  • Treating CCPA as GDPR-lite — opt-out vs opt-in inverts the default. Pre-checking marketing in CA is fine; pre-sharing with ad networks is not.
  • 45-day SLA without queue — manual response to high-volume DSARs misses windows. Build verifiable-request intake → identity-proof → workflow with timer.
  • SPI inference without limit-use link — using ZIP+name to infer ethnicity for ad targeting triggers SPI even if the source field wasn't SPI. Inference creates SPI.
  • Ignoring authorized agents — CCPA allows power-of-attorney delegation; refusing agent-submitted requests is an automatic violation.

Handoff

  • To Crypt: encryption design for SPI at rest/in transit (CCPA "reasonable security" defense). Required for credentials, financial, geolocation, biometric.
  • To Canon: CCPA standard checklist cross-validation against ISO/IEC 27701 PIMS controls; gap matrix.
  • To Canon: Privacy Policy canon review — disclose categories, purposes, retention, sale/share, SPI, rights. Tokushoho-style equivalents for CA-targeted commerce.
  • To Canon[regulatory]: SOC 2 / ISO 27001 mapping for CCPA risk-assessment + cybersecurity-audit controls; evidence package.
  • To Builder: Implement opt-out endpoint, GPC parser, SPI limit-use toggle, ADMT pre-use notice, deletion-propagation queue.
  • To Schema: tag SPI fields with spi_category annotation; persist gpc_honored_at, opt_out_status, limit_spi_status per consumer.
  • To Gateway: header-level GPC detection middleware; consent-state propagation to downstream services.

Source: SKILL.md on GitHub

No alerts13d4 checks · Risk SAFE
  • Gen Agent Trust Hub13d

    The skill is a specialized tool for privacy engineering and regulatory compliance. The analysis identifies a risk of indirect prompt injection because the skill's primary function is to ingest and analyze potentially untrusted external data sources, such as source code, application logs, and database schemas. While the skill mandates PII redaction, it lacks explicit boundary markers to prevent the agent from inadvertently following instructions embedded within the data it scans. No other malicious behaviors or data exfiltration attempts were detected.

  • Socket13d

    No alerts

  • Snyk13d

    Risk: LOW · No issues

  • ZeroLeaks5mo

    Score: 93/100 · 2 sections analyzed

Signed by skilld at 35ffd55. 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/cloak