Sushegaad/Claude-Skills-Governance-Risk-and-Compliance / plugins/eu-ai-act/skills/eu-ai-act
eu-ai-act
VerifiedEU AI Act (Regulation (EU) 2024/1689) compliance advisor - risk classification across all four tiers, all 9 prohibited practices (Art. 5, including the nudification/CSAM prohibition from Dec 2, 2026), all 8 Annex III high-risk use case areas, provider and deployer obligations (Arts. 9–17, 26), GPAI model obligations including the July 2025 Code of Practice (Arts. 51–55), conformity assessment and CE marking (Arts. 43–48), EU AI database registration, Art. 50 transparency (chatbots, synthetic media, AI-generated content), governance (AI Office, AI Board), penalties (Art. 99), confirmed phase-in timeline (Digital Omnibus, Reg. (EU) 2026/1744, in force July 27, 2026: Annex III deferred to Dec 2, 2027; Annex I to Aug 2, 2028), and cross-framework mapping to ISO 42001, NIST AI RMF, and GDPR. Use for any EU AI regulation, AI system classification, or AI compliance question. Current as of November 2026. GPAI enforcement active: Art. 91 information requests reported since Aug 2026 (procedures: Reg. (EU) 2026/1755).
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 30.1 KB |
| references/gpai-governance.md | markdown | 19.8 KB |
| references/obligations-high-risk.md | markdown | 15.4 KB |
| references/risk-classification.md | markdown | 14.3 KB |
references/gpai-governance.md
View on GitHub# EU AI Act - GPAI, Governance, and Cross-Framework Reference
## GPAI Model Obligations
### Art. 3(63) - What Qualifies as a GPAI Model
A model qualifies as GPAI when **all** of the following apply:
- Trained with **large amounts of data** using **self-supervision at scale**
- Displays **significant generality** - competently performs a wide range of distinct tasks
- Can be used for multiple **different purposes** regardless of how placed on market
- **Excludes:** Models in pre-release R&D phase not yet placed on market
**Key implication:** Many foundation models (large language models, multimodal models) are GPAI models. A system built on such a model may be a "GPAI system" (Art. 3(64)).
---
### Art. 53 - Universal GPAI Provider Obligations (applies from 2 August 2025)
ALL GPAI model providers (regardless of size or systemic risk status) must:
**1. Technical Documentation (Annex XI content):**
- General description: intended purpose, types of tasks, interaction modes
- Development process: data sources, data processing and filtering, training techniques, compute used
- Testing and evaluation: benchmarks, evaluations, safety testing results
- Known limitations and foreseeable risks
- Responsible disclosure practices; contact information for reporting issues
- Keep up-to-date; provide to AI Office and national authorities on request
**2. Downstream Provider Information (Annex XII content):**
Document and make available to downstream AI system providers who integrate the GPAI model:
- Capabilities and limitations enabling compliance with their own obligations
- Integration instructions
- Security measures
- Evaluation results
- **Note:** Intellectual property protections are permitted - Annex XII information may be provided under reasonable commercial terms
**3. Copyright Compliance Policy (EU Directive 2019/790 Text and Data Mining):**
- Implement a policy to comply with EU copyright and related rights law
- Specifically: mechanisms to respect text and data mining opt-outs reserved by rights-holders under Art. 4(3) of Directive 2019/790
- Must be technically enforceable and publicly described
**4. Public Training Content Summary:**
- Draw up and make publicly available a sufficiently detailed summary of training content
- Must follow Commission-provided template
- Purpose: enable downstream providers and the public to understand what data was used
**Open-source / free-license exception:**
GPAI model providers releasing under open-source/free licenses need only comply with obligations **3 and 4** (copyright policy and training summary) - **UNLESS** the model is later designated as having systemic risk, in which case full Art. 53 + Art. 55 obligations apply.
---
### Art. 51 - Systemic Risk Classification
**GPAI classification threshold (Commission guidelines, July 2025):**
Models trained with cumulative compute ≥ **10²³ FLOPs** are subject to GPAI obligations under Arts. 53–55.
**Automatic presumption of systemic risk when:**
Training compute exceeds **10²⁵ FLOPs** (floating point operations, cumulative across all training runs)
**Commission designation (independent of FLOPs threshold):**
Commission may also designate a model as having systemic risk based on:
- High-impact capabilities identified through model evaluation
- Broad reach across the EU
- Negative effects on public health, safety, public security, fundamental rights
- Other criteria in Annex XIII (including parameter count, domain specificity, user base)
**Provider notification obligation:**
Providers that know or should reasonably know their model meets or is approaching the threshold must:
- Notify the **AI Office** within **2 weeks** of reaching the threshold
- Notification triggers systemic risk assessment dialogue
**Dynamic threshold:**
Commission may amend the 10²⁵ FLOPs threshold via delegated acts as AI capabilities evolve.
**Rebuttable presumption:**
Providers may present evidence demonstrating systemic risk does not apply despite meeting the FLOPs threshold.
---
### Art. 55 - Systemic Risk GPAI Provider Additional Obligations (applies from 2 August 2025)
In addition to Art. 53 universal obligations, providers of systemic-risk GPAI models must:
**1. Model Evaluation:**
- Conduct model evaluation according to standardised protocols and state-of-the-art tools
- Includes adversarial testing (red-teaming) to identify and mitigate systemic risks
- May be conducted in cooperation with the AI Office
**2. Risk Assessment and Mitigation:**
- Assess and mitigate possible systemic risks arising from development, market placement, or deployment
- Systemic risks include: actual or foreseeable negative effects on public health, safety, public security, or fundamental rights at Union level, including serious societal disturbances
**3. Serious Incident Reporting:**
- Track, document, and promptly report any serious incidents and possible corrective actions
- Report to: **AI Office** and relevant **national competent authorities**
- Report without undue delay upon becoming aware
**4. Cybersecurity:**
- Ensure adequate cybersecurity protection for the model, its physical infrastructure, and the supply chain
- Includes protection against model theft, unauthorized access, adversarial attacks on model weights
**Compliance pathways for GPAI obligations:**
1. **GPAI Code of Practice** (primary pathway) - Final CoP published 10 July 2025, endorsed by Commission and AI Board 1 August 2025. Operates from 2 August 2025. Three chapters: (i) Transparency, (ii) Copyright, (iii) Safety and Security (systemic risk only). Non-signatories must demonstrate compliance by alternative means and explain their approach to the AI Office.
2. **Harmonised standards** - when published under Commission mandate; compliance gives presumption of conformity. **EN 18286:2026** (AI quality management, supporting Art. 17) became the first AI Act harmonised standard cited in the OJEU, announced October 15, 2026. The broader standardisation request deadline was revised to February 2027; prEN 18228 (risk management) is in final approval.
3. **Alternative adequate means** - provider demonstrates compliance through other methods acceptable to Commission.
**GPAI Code of Practice - Key obligations by chapter:**
- **Transparency chapter:** Model Documentation Form covering technical properties, training data, energy consumption, intended use; retained for 10 years. Downstream providers can request additional information within 14 days.
- **Copyright chapter:** Written copyright policy; compliance with robots.txt and technical opt-out protocols; rightsholder complaints mechanism; prohibition of infringing uses in acceptable use policies.
- **Safety and Security chapter (systemic risk only):** Safety and Security Framework established within **4 weeks of notification to the Commission** and **at least 2 weeks before market placement**; Safety and Security Model Reports submitted to AI Office before market placement; adversarial testing (red-teaming); serious incident reporting with staggered timelines; cybersecurity measures.
**Legacy GPAI models** (placed on market before 2 August 2025): deadline to comply is **2 August 2027**.
**Confirmed CoP signatories:** 22 full signatories per the Commission page updated October 7, 2026, including Amazon, Anthropic, Google, IBM, Microsoft, Mistral AI, and OpenAI. Note: xAI signed only the Safety and Security chapter, committing to demonstrate compliance with Transparency and Copyright by alternative means.
A **Signatory Taskforce**, chaired by the AI Office, facilitates coherent CoP implementation.
---
## Governance Structure
### AI Office (Art. 64)
- Established within the **European Commission** (not an independent agency)
- Primary responsibility for: GPAI model oversight, monitoring, compliance assessment
- Conducts GPAI model evaluations - both compliance assessments and systemic risk investigations
- May request access to model weights, source code, training procedures, technical documentation
- Receives complaints from downstream providers against upstream GPAI providers
- Provides secretariat support for the AI Board
- Key contact point for: systemic risk notifications, Codes of Practice development, serious incident reports from GPAI providers
- Issues non-binding guidance; makes recommendations to the Commission
- Full enforcement powers over GPAI providers (information requests, model evaluations, fines) active since **2 August 2026**; procedural rules for Commission proceedings are set by **Implementing Regulation (EU) 2026/1755** (OJ July 21, 2026, in force August 10, 2026): model evaluations including API, source code, and model access; appointment and independence of experts; fine proceedings with preliminary findings and a right to be heard; access to file; five-year limitation period
- Enforcement in practice: first formal Art. 91 requests for information to GPAI providers reported in late August 2026; a further wave was announced October 14, 2026 covering security practices, external evaluation, and post-market monitoring; recipients have not been officially named
**Key guidance and actions published by the AI Office (chronological):**
- **4 Feb 2025:** Guidelines on prohibited AI practices under Art. 5 - with practical examples for all 8 categories
- **6 Feb 2025:** Guidelines on the definition of an AI system (Art. 3(1)) - clarifying what falls within the regulation's scope
- **July 2025:** Final GPAI Code of Practice published; GPAI scope guidelines published; Commission training data summary template published (for Art. 53(1)(d))
- **October 2025:** AI Act Whistleblower Tool launched - secure channel for reporting suspected violations
- **December 2025:** First draft of Code of Practice on Marking and Labelling of AI-generated Content
- **5 March 2026:** Second draft of the AI-generated content marking CoP
- **8 May 2026:** Draft guidelines on Art. 50 transparency obligations opened for consultation; three studies on Art. 5 prohibited practices published; three studies on technical marking/detection of AI-generated content published
- **19 May 2026:** Draft guidelines on high-risk AI system classification opened for public consultation
- **21 July 2026:** Implementing Regulation (EU) 2026/1755 published in the OJ (in force 10 August 2026): procedural rules for Commission proceedings against GPAI providers
- **Late August 2026:** First formal Art. 91 requests for information to GPAI providers reported (recipients not officially named)
- **12 October 2026:** Consultation opened on a draft implementing act for penalty-proceeding procedural rules (closing 9 November 2026)
- **14 October 2026:** Further wave of Art. 91 information requests announced, covering security practices, external evaluation, and post-market monitoring
- **15 October 2026:** EN 18286:2026 announced as the first AI Act harmonised standard cited in the OJEU (Art. 17 quality management)
- Note: the Art. 73 serious-incident reporting guidance remains a DRAFT (September 2025); no final version as of November 2026
### European Artificial Intelligence Board - AI Board (Art. 65)
**Composition:**
- One representative per EU Member State (3-year renewable mandate)
- European Data Protection Supervisor: permanent observer
- AI Office representative: attends, no voting rights
- Other national/Union authorities: invited case-by-case
**Governance:**
- Elects Chair from Member State representatives
- Adopts rules of procedure by two-thirds majority
- Two standing sub-groups: (i) market surveillance authorities; (ii) notifying authorities
**Core tasks:**
- Advise and assist Commission and Member States on consistent AI Act implementation
- Issue opinions, recommendations, and guidance on matters affecting multiple Member States
- Coordinate and facilitate information exchange between national authorities
- Facilitate development of harmonised technical standards
- Provide opinions on common specifications
### National Competent Authorities
Each Member State designates one or more national competent authorities with powers covering:
- **Market surveillance** (Art. 74): access to documentation, testing datasets, source code
- **Conformity assessment oversight**: notified body designation and monitoring
- **Enforcement and investigation**: power to require corrective action, withdrawal, recalls, fines
- **Annual reporting** to Commission on prohibited practices and enforcement actions
**Sector-specific authorities:**
- Financial supervisors for AI in banking and insurance
- Data protection authorities for law enforcement AI applications
- European Data Protection Supervisor for EU institutions' AI systems (Art. 77)
### Advisory Forum (Art. 67)
- Stakeholder body advising the Commission and the AI Board
- Balanced membership drawn from industry (including SMEs and start-ups), civil society, and academia
- Provides technical expertise and stakeholder input for the Board's and the Commission's tasks
### Scientific Panel (Art. 68)
- Independent expert body for GPAI matters
- Can alert AI Office about high-impact GPAI capabilities that may qualify for systemic risk designation
- Provides technical opinions to AI Office and Commission
- Issues views on model evaluation methodologies and capabilities assessment
### AI Regulatory Sandboxes (Art. 57)
- At least one operational per Member State by **2 August 2027** (deadline extended from 2 Aug 2026 by AI Omnibus)
- Enable controlled testing of innovative AI systems under regulatory supervision before market placement
- Participants: reduced administrative burden; authorities: supervisory insight
- Participation does not provide automatic compliance certification
---
## Post-Market Monitoring and Incident Reporting
### Art. 72 - Post-Market Monitoring System
**Applies to:** All high-risk AI system providers.
**Requirements:**
- Establish and document a post-market monitoring system proportionate to the nature of the AI technology and the risk
- Continuously collect, document, and analyze relevant performance data from deployed systems throughout the operational lifetime
- Monitor for risks to health, safety, and fundamental rights
- Inform risk management system updates (Art. 9)
- Must form part of technical documentation (Art. 11)
**Commission template:** Commission must publish a standardised template for post-market monitoring plans by 2 February 2026.
**Integration:** Existing post-market monitoring systems under sectoral legislation (medical devices, machinery, etc.) may be adapted and integrated to avoid duplication.
### Art. 73 - Serious Incident Reporting
**What constitutes a serious incident:**
Any incident or malfunctioning of a high-risk AI system that directly or indirectly leads or could lead to:
- Death of a person or serious damage to health of a person
- Serious and irreversible disruption to management/operation of critical infrastructure
- Infringement of obligations under Union law protecting fundamental rights
- Serious damage to property or the environment
**Who reports:**
- **Providers:** Report to market surveillance authority of Member State where incident occurred; within timeframe proportionate to severity (for death-related incidents, Art. 73(4): immediately upon establishing a causal link between the AI system and the death, and no later than 10 days after awareness; within 15 days for other serious incidents; within 2 days for widespread safety threats)
- **Deployers:** Immediately notify provider, importer/distributor, AND market surveillance authority
- **GPAI systemic risk providers:** Report to AI Office and relevant national authorities
---
## Cross-Framework Mapping
### ISO 42001:2023 (AI Management System)
ISO 42001 is the primary AI management system standard complementing AI Act compliance.
| AI Act Article | ISO 42001 Element |
|---------------|-------------------|
| Art. 9 (Risk Management) | Clause 6.1 (Actions to address risks and opportunities), A.5.2 (AI system impact assessment process) |
| Art. 10 (Data Governance) | Clause 8 (Operation), A.7 (Data for AI systems: A.7.2 to A.7.6) |
| Art. 13 (Transparency) | A.8.2 (System documentation and information for users), A.6.2.7 (AI system technical documentation) |
| Art. 14 (Human Oversight) | A.9.2 (Processes for responsible use of AI systems), A.6.1.3 (Processes for responsible AI system design and development) |
| Art. 15 (Accuracy/Robustness) | A.6.2.4 (AI system verification and validation), A.6.2.6 (AI system operation and monitoring) |
| Art. 17 (QMS) | Clause 4–10 PDCA management system structure |
| Art. 72 (Post-Market Monitoring) | Clause 9.1 (Monitoring, measurement, analysis and evaluation), A.6.2.6 (AI system operation and monitoring) |
| Art. 73 (Incident Reporting) | A.8.4 (Communication of incidents), Clause 10.2 (Nonconformity and corrective action) |
**Practical guidance:** ISO 42001 certification can support demonstration of Art. 17 QMS compliance. ISO 42001's Statement of Applicability (SoA) process may be adapted for mapping conformity with AI Act requirements.
### NIST AI RMF (AI Risk Management Framework 1.0)
| AI Act Obligation | NIST AI RMF Function/Category |
|------------------|-------------------------------|
| Art. 17 QMS | GOVERN 1 (Policies, processes, accountability) |
| Art. 6 classification | MAP 1 (Context establishment), MAP 2 (Risk identification) |
| Art. 9 risk management | MAP 3, MEASURE 1–2 |
| Art. 10 data governance | MEASURE 2.5, 2.6 (Data and bias evaluation) |
| Art. 15 accuracy/robustness | MEASURE 2.1–2.4 (Testing, evaluation, red-teaming) |
| Art. 14 human oversight | MANAGE 2 (Human-AI interaction), GOVERN 4 |
| Art. 72 post-market monitoring | MANAGE 3, MANAGE 4 (Post-deployment operations) |
| Art. 73 incident reporting | MANAGE 4.1–4.2 (Incident response) |
### GDPR Alignment
The AI Act operates concurrently with GDPR - both apply when AI systems process personal data.
| Intersection | AI Act | GDPR |
|-------------|--------|------|
| Data governance | Art. 10 | Art. 5 (data quality principles), Art. 25 (data protection by design) |
| DPIA requirement | Art. 26 (deployer obligation reference) | Art. 35 |
| Special category data | Art. 10(5) (bias detection) | Art. 9 |
| Profiling definition | Art. 3(52) (matches precisely) | Art. 4(4) |
| Transparency to users | Art. 50 | Art. 13/14 (information notices) |
| Fundamental rights impact | Art. 27 (FRIA for public authorities) | Art. 35 DPIA |
| RBI system authorisation | Art. 5(1)(h) | Art. 9 + law enforcement exemptions |
| Supervisory authorities | Art. 77 (DPAs have access rights) | Supervisory authority under Chapter VI |
| Enforcement | Art. 99 (AI Act fines) | Art. 83 (GDPR fines up to 4% global turnover) |
**Key distinction:** GDPR governs personal data processing; the AI Act governs AI system development and deployment. Many AI systems trigger obligations under both - operators need dual compliance programmes.
---
## Penalties Reference - Art. 99
| Violation Type | Maximum Fine |
|----------------|--------------|
| **Prohibited AI practices** (Art. 5 violations) | €35,000,000 or **7%** of total worldwide annual turnover (higher applies) |
| **Provider/deployer/notified body violations** (Arts. 16, 22, 23, 24, 26, 31, 33, 34, 50) | €15,000,000 or **3%** of total worldwide annual turnover (higher applies) |
| **Incorrect/misleading information** to notified bodies or competent authorities | €7,500,000 or **1%** of total worldwide annual turnover (higher applies) |
**SME / startup / SMC reduction:** For SMEs, startups, and - per the AI Omnibus - **small mid-cap companies (SMCs)**, the lower of the fixed amount or percentage applies (Art. 99(6)).
**Proportionality factors:** Nature, gravity, and duration; prior infringements; company size and market share; intentionality; cooperation level; mitigation measures taken; actual damage caused.
**Member State reporting:** Annual reporting obligation to Commission on fines issued.
**GPAI enforcement:** AI Office enforces GPAI obligations. For systemic risk GPAI models, Commission itself may conduct investigations and impose fines directly.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| writing-skillsobra/superpowersUse when creating new skills, editing existing skills, or verifying skills work before deployment | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| xlsxanthropics/skillsUse this skill any time a spreadsheet file is the primary input or output. This means any task where the user wants to: open, read, edit, or fix an existing .xlsx, .xlsm, .xltx, .csv, or .tsv file (e.g., adding columns, computing formulas, formatting, charting, cleaning messy data); create a new spreadsheet from scratch or from other data sources; or convert between tabular file formats. Trigger especially when the user references a spreadsheet file by name or path — even casually (like "the xlsx in my downloads") — and wants something done to it or produced from it. Also trigger for cleaning or restructuring messy tabular data files (malformed rows, misplaced headers, junk data) into proper spreadsheets. The deliverable must be a spreadsheet file. Do NOT trigger when the primary deliverable is a Word document, HTML report, standalone Python script, database pipeline, or Google Sheets API integration, even if tabular data is involved. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| shipping-and-launchaddyosmani/agent-skillsPrepares production launches. Use when preparing to deploy to production, or when asking what needs to be in place before shipping. Use when you need a pre-launch checklist, when setting up monitoring, when planning a staged rollout, or when you need a rollback strategy. | Stars: 119k | Updated | Claude CodeCodex | Verified | Install |
| childhoodthedotmack/claude-memBabysit a shipped release after merge, publication, or deployment. Verify published packages, production behavior, installed plugins, and late CI or user reports until the observation window is clean; use babysit for PR readiness before release. | Stars: 118k | Updated | Claude CodeCodex | Verified | Install |