+Article 55 Obligations of providers of general-purpose AI models with systemic risk
|
Article 55 Obligations of providers of general-purpose AI models with systemic risk
Article 55
Obligations of providers of general-purpose AI models with systemic risk
1. In addition to the obligations listed in Articles 53 and 54, providers of general-purpose AI models with systemic risk shall:
|
(a)
|
perform model evaluation in accordance with standardised protocols and tools reflecting the state of the art, including conducting and documenting adversarial testing of the model with a view to identifying and mitigating systemic risks;
|
|
(b)
|
assess and mitigate possible systemic risks at Union level, including their sources, that may stem from the development, the placing on the market, or the use of general-purpose AI models with systemic risk;
|
|
(c)
|
keep track of, document, and report, without undue delay, to the AI Office and, as appropriate, to national competent authorities, relevant information about serious incidents and possible corrective measures to address them;
|
|
(d)
|
ensure an adequate level of cybersecurity protection for the general-purpose AI model with systemic risk and the physical infrastructure of the model.
|
2. Providers of general-purpose AI models with systemic risk may rely on codes of practice within the meaning of Article 56 to demonstrate compliance with the obligations set out in paragraph 1 of this Article, until a harmonised standard is published. Compliance with European harmonised standards grants providers the presumption of conformity to the extent that those standards cover those obligations. Providers of general-purpose AI models with systemic risks who do not adhere to an approved code of practice or do not comply with a European harmonised standard shall demonstrate alternative adequate means of compliance for assessment by the Commission.
3. Any information or documentation obtained pursuant to this Article, including trade secrets, shall be treated in accordance with the confidentiality obligations set out in Article 78.
1. Overview
1.1 References
1.2 Identified Requirements
1.3 Related Standards
2. Identified Requirements
Requirements
| Source |
Requirement |
3. Related Standards
Standards
| Source |
Requirement |
|
SCF
|
Adequate Protections For AI & Autonomous Technologies
Description
Mechanisms exist to ensure Artificial Intelligence (AI) and Autonomous Technologies (AAT) include reasonable security, compliance and resilience protections that are commensurate with assessed risks and threats.
Possible Solutions & Considerations
Micro-Small Business (<10 staff) / BLS Firm Size Classes 1-2
∙ Contractual security requirements in AI vendor agreements
Small Business (10-49 staff) / BLS Firm Size Classes 3-4
∙ Security assessment of AI tools before adoption
∙ Vendor security questionnaires
Medium Business (50-249 staff) / BLS Firm Size Classes 5-6
∙ AI security requirements checklist
∙ Third-party risk assessment for AI vendors
Large Business (250-999 staff) / BLS Firm Size Classes 7-8
∙ Formal AI security standards
∙ Security review in AI procurement process
∙ Ongoing vendor monitoring
Enterprise (> 1,000 staff) / BLS Firm Size Class 9
∙ Enterprise AI security requirements framework
∙ Automated TPRM platform
∙ Continuous AI vendor monitoring
SCR-CMM
Level 0 Not Performed
Practices are non-existent, based on the inability to demonstrate an implemented and operational capability. A reasonable person would conclude the control is not being performed.
Level 1 Performed Informally
SCR-CMM Level 1 criteria definitions are not available for this control:
▪ A reasonable person would conclude this control requires a structured process.
▪ At this level of maturity, the "ad hoc" nature of performing a capability informally would indicate the intent of the control is not met due to a lack of consistency and formality.
Level 2 Planned Tracked
Artificial Intelligence and Autonomous Technology (AAT) capabilities are requirements-driven, but are not standardized across the entity (e.g., local/regional level consistency). Capability criteria associated with this control reasonably expect the following criteria to exist:
▪ Policies and standards associated with AAT domain capabilities are formally documented and centrally-managed by the entity.
▪ Standardized Operating Procedures (SOP) associated with AAT domain capabilities are documented and maintained by process owners.
▪ IT and/or cybersecurity personnel work with business stakeholders and process owners to appropriately scope and reasonably implement cybersecurity and data protection controls associated with AAT domain capabilities to address applicable statutory, regulatory and/or contractual requirements for Technology Assets, Applications, Services and/or Data (TAASD).
▪ Artificial Intelligence (AI)-related controls are primarily administrative and preventative in nature (e.g., policies, standards, procedures & guidelines).
▪ Asset management may be a defined function (e.g., team or department) or assigned as an additional duty to existing IT and/or cybersecurity personnel.
Level 3 Well Defined
Artificial Intelligence and Autonomous Technology (AAT) capabilities are standardized across the entity for applicability to People, Processes, Technologies, Data and/or Facilities (PPTDF) to ensure consistency for Technology Assets, Applications, Services and/or Data (TAASD). Capability criteria associated with this control reasonably expect the following criteria to exist:
▪ Policies and standards associated with AAT domain capabilities are formally documented and centrally-managed by the entity's Governance, Risk & Compliance (GRC) team, or similar function.
▪ Standardized Operating Procedures (SOP) associated with AAT domain capabilities are well-documented and kept current by process owners.
▪ An Artificial Intelligence Governance (AIG) team, or similar function, is appropriately staffed and supported to implement and maintain AAT domain capabilities.
▪ Technology is leveraged to enhance the efficiency and accuracy of AI governance, risk management and compliance operations (e.g., dedicated AI governance platform).
▪ The entity's Governance, Risk & Compliance (GRC) team, or similar function, works with business stakeholders and process owners to appropriately scope and reasonably implement cybersecurity and data protection controls associated with AAT domain capabilities to address Minimum Compliance Requirements (MCR) (e.g., applicable statutory, regulatory and/or contractual requirements) and Discretionary Security Requirements (DSR) (e.g., entity-required controls).
▪ An implemented and operational capability exists to ensure AAT include reasonable security, compliance and resilience protections that are commensurate with assessed risks and threats.
Level 4 Quantitatively Controlled
Utilize SCR-CMM Level 3 criteria definitions:
▪ There are no defined Level 4 criteria, since it is reasonable to assume a quantitatively-controlled process is not necessary to operationalize this control.
▪ While it may be possible to develop “metrics-driven” capabilities for this control, the criteria would be organization-specific to define.
Level 5 Continuously Improving
Utilize SCR-CMM Level 3 or Level 4 (if available) criteria definitions:
▪ There are no defined Level 5 criteria, since it is reasonable to assume a continuously-improving process is not necessary to operationalize this control.
▪ Level 5 capabilities should be considered “world-class” where the control builds on Level 4 capabilities, but are continuously improving through Artificial Intelligence (AI) and/or Machine Learning (ML) technologies.
▪ While it may be possible to develop responsive capabilities for this control through the use of AI and/or ML technologies, the criteria would be organization-specific to define.
|
|
SCF
|
Artificial Intelligence Test, Evaluation, Validation & Verification (AI TEVV)
Description
Mechanisms exist to implement Artificial Intelligence Test, Evaluation, Validation & Verification (AI TEVV) practices to enable Artificial Intelligence (AI) and Autonomous Technologies (AAT)-related security, resilience and compliance-related conformity testing throughout the lifecycle of the AAT.
Possible Solutions & Considerations
Micro-Small Business (<10 staff) / BLS Firm Size Classes 1-2
∙ Information Assurance (IA) Program
∙ AI TEVV checklist for AI tools (test accuracy, validate outputs, verify security)
∙ AI governance program
Small Business (10-49 staff) / BLS Firm Size Classes 3-4
∙ Information Assurance (IA) Program
∙ Formal AI TEVV process for AI systems
∙ AI governance program
Medium Business (50-249 staff) / BLS Firm Size Classes 5-6
∙ Information Assurance (IA) Program
∙ Formal AI TEVV framework aligned to NIST AI RMF Measure function
∙ AI testing tools (e.g., IBM OpenScale, Great Expectations)
∙ AI governance program
Large Business (250-999 staff) / BLS Firm Size Classes 7-8
∙ Information Assurance (IA) Program
∙ Enterprise AI TEVV program
∙ NIST AI RMF Measure function
∙ Third-party AI testing for high-risk systems
∙ AI testing integrated with CI/CD pipelines
Enterprise (> 1,000 staff) / BLS Firm Size Class 9
∙ Information Assurance (IA) Program
∙ Enterprise AI TEVV program (NIST AI RMF Measure function)
∙ Independent AI testing and evaluation for high-risk systems
∙ AI TEVV integrated with MLOps and CI/CD
∙ EU AI Act conformity assessment (if applicable)
SCR-CMM
Level 0 Not Performed
Practices are non-existent, based on the inability to demonstrate an implemented and operational capability. A reasonable person would conclude the control is not being performed.
Level 1 Performed Informally
SCR-CMM Level 1 criteria definitions are not available for this control:
▪ A reasonable person would conclude this control requires a structured process.
▪ At this level of maturity, the "ad hoc" nature of performing a capability informally would indicate the intent of the control is not met due to a lack of consistency and formality.
Level 2 Planned Tracked
SCR-CMM Level 2 criteria definitions are not available for this control:
▪ A reasonable person would conclude a well-defined and standardized process is required.
▪ At this level of maturity, the “requirements-driven” nature of performing the control is focused on a localized and/or regionalized implementation, not uniform and consistent across the organization.
▪ Requirements are narrowly scoped for applicability and are primarily derived from compliance obligations (e.g., laws, regulations and contracts).
Level 3 Well Defined
Artificial Intelligence and Autonomous Technology (AAT) capabilities are standardized across the entity for applicability to People, Processes, Technologies, Data and/or Facilities (PPTDF) to ensure consistency for Technology Assets, Applications, Services and/or Data (TAASD). Capability criteria associated with this control reasonably expect the following criteria to exist:
▪ Policies and standards associated with AAT domain capabilities are formally documented and centrally-managed by the entity's Governance, Risk & Compliance (GRC) team, or similar function.
▪ Standardized Operating Procedures (SOP) associated with AAT domain capabilities are well-documented and kept current by process owners.
▪ An Artificial Intelligence Governance (AIG) team, or similar function, is appropriately staffed and supported to implement and maintain AAT domain capabilities.
▪ Technology is leveraged to enhance the efficiency and accuracy of AI governance, risk management and compliance operations (e.g., dedicated AI governance platform).
▪ The entity's Governance, Risk & Compliance (GRC) team, or similar function, works with business stakeholders and process owners to appropriately scope and reasonably implement cybersecurity and data protection controls associated with AAT domain capabilities to address Minimum Compliance Requirements (MCR) (e.g., applicable statutory, regulatory and/or contractual requirements) and Discretionary Security Requirements (DSR) (e.g., entity-required controls).
▪ An implemented and operational capability exists to implement Artificial Intelligence Test, Evaluation, Validation & Verification (AI TEVV) practices to enable AAT-related security, resilience and compliance-related conformity testing throughout the lifecycle of the AAT.
Level 4 Quantitatively Controlled
Utilize SCR-CMM Level 3 criteria definitions:
▪ There are no defined Level 4 criteria, since it is reasonable to assume a quantitatively-controlled process is not necessary to operationalize this control.
▪ While it may be possible to develop “metrics-driven” capabilities for this control, the criteria would be organization-specific to define.
Level 5 Continuously Improving
Utilize SCR-CMM Level 3 or Level 4 (if available) criteria definitions:
▪ There are no defined Level 5 criteria, since it is reasonable to assume a continuously-improving process is not necessary to operationalize this control.
▪ Level 5 capabilities should be considered “world-class” where the control builds on Level 4 capabilities, but are continuously improving through Artificial Intelligence (AI) and/or Machine Learning (ML) technologies.
▪ While it may be possible to develop responsive capabilities for this control through the use of AI and/or ML technologies, the criteria would be organization-specific to define.
|
|
SCF
|
Serious Incident Reporting For AI & Autonomous Technologies
Description
Mechanisms exist to report any serious incident involving operational Artificial Intelligence (AI) and Autonomous Technologies (AAT) to relevant authorities as to when and where the serious incident occurred, in accordance with mandated reporting timelines.
Possible Solutions & Considerations
Micro-Small Business (<10 staff) / BLS Firm Size Classes 1-2
∙ Report serious AI incidents to relevant authorities as required
Small Business (10-49 staff) / BLS Firm Size Classes 3-4
∙ Incident reporting procedure for serious AI incidents
Medium Business (50-249 staff) / BLS Firm Size Classes 5-6
∙ Formal serious AI incident reporting process
∙ Regulatory reporting contacts
Large Business (250-999 staff) / BLS Firm Size Classes 7-8
∙ AI incident response plan with regulatory reporting requirements
∙ Legal team involvement
Enterprise (> 1,000 staff) / BLS Firm Size Class 9
∙ Enterprise AI incident response program
∙ Automated regulatory notification workflows
∙ Legal and compliance integration
SCR-CMM
Level 0 Not Performed
Practices are non-existent, based on the inability to demonstrate an implemented and operational capability. A reasonable person would conclude the control is not being performed.
Level 1 Performed Informally
Artificial Intelligence and Autonomous Technology (AAT) domain capabilities are ad hoc and inconsistent. Capability criteria associated with this control may include:
▪ Policies, standards & procedures associated with AAT domain capabilities provide limited coverage due to the depth and breadth of the existing documentation.
▪ AAT-related processes are expected to follow the organization's existing processes (e.g., incident response, asset management, change control, risk assessments, etc.).
▪ No formal Governance, Risk & Compliance (GRC) team exists to provide AAT oversight, where the Chief Information Officer (CIO), or similar function, governs technology decisions what is acceptable for AAT within the organization.
Level 2 Planned Tracked
Artificial Intelligence and Autonomous Technology (AAT) capabilities are requirements-driven, but are not standardized across the entity (e.g., local/regional level consistency). Capability criteria associated with this control reasonably expect the following criteria to exist:
▪ Policies and standards associated with AAT domain capabilities are formally documented and centrally-managed by the entity.
▪ Standardized Operating Procedures (SOP) associated with AAT domain capabilities are documented and maintained by process owners.
▪ IT and/or cybersecurity personnel work with business stakeholders and process owners to appropriately scope and reasonably implement cybersecurity and data protection controls associated with AAT domain capabilities to address applicable statutory, regulatory and/or contractual requirements for Technology Assets, Applications, Services and/or Data (TAASD).
▪ Artificial Intelligence (AI)-related controls are primarily administrative and preventative in nature (e.g., policies, standards, procedures & guidelines).
▪ Asset management may be a defined function (e.g., team or department) or assigned as an additional duty to existing IT and/or cybersecurity personnel.
Level 3 Well Defined
Artificial Intelligence and Autonomous Technology (AAT) capabilities are standardized across the entity for applicability to People, Processes, Technologies, Data and/or Facilities (PPTDF) to ensure consistency for Technology Assets, Applications, Services and/or Data (TAASD). Capability criteria associated with this control reasonably expect the following criteria to exist:
▪ Policies and standards associated with AAT domain capabilities are formally documented and centrally-managed by the entity's Governance, Risk & Compliance (GRC) team, or similar function.
▪ Standardized Operating Procedures (SOP) associated with AAT domain capabilities are well-documented and kept current by process owners.
▪ An Artificial Intelligence Governance (AIG) team, or similar function, is appropriately staffed and supported to implement and maintain AAT domain capabilities.
▪ Technology is leveraged to enhance the efficiency and accuracy of AI governance, risk management and compliance operations (e.g., dedicated AI governance platform).
▪ The entity's Governance, Risk & Compliance (GRC) team, or similar function, works with business stakeholders and process owners to appropriately scope and reasonably implement cybersecurity and data protection controls associated with AAT domain capabilities to address Minimum Compliance Requirements (MCR) (e.g., applicable statutory, regulatory and/or contractual requirements) and Discretionary Security Requirements (DSR) (e.g., entity-required controls).
▪ An implemented and operational capability exists to report any serious incident involving operational AAT to relevant authorities as to when and where the serious incident occurred, in accordance with mandated reporting timelines.
Level 4 Quantitatively Controlled
Utilize SCR-CMM Level 3 criteria definitions:
▪ There are no defined Level 4 criteria, since it is reasonable to assume a quantitatively-controlled process is not necessary to operationalize this control.
▪ While it may be possible to develop “metrics-driven” capabilities for this control, the criteria would be organization-specific to define.
Level 5 Continuously Improving
Utilize SCR-CMM Level 3 or Level 4 (if available) criteria definitions:
▪ There are no defined Level 5 criteria, since it is reasonable to assume a continuously-improving process is not necessary to operationalize this control.
▪ Level 5 capabilities should be considered “world-class” where the control builds on Level 4 capabilities, but are continuously improving through Artificial Intelligence (AI) and/or Machine Learning (ML) technologies.
▪ While it may be possible to develop responsive capabilities for this control through the use of AI and/or ML technologies, the criteria would be organization-specific to define.
|
|