Why the Act matters for CISOs
The EU AI Act is not only a legal-policy exercise. Its requirements for accuracy, robustness, cybersecurity, logging, and oversight can give CISOs a substantial implementation role, alongside product, risk, legal, and other accountable teams.
Unlike GDPR, which focuses on data protection policies and procedures, the AI Act demands demonstrable technical controls. Article 15 explicitly requires “appropriate levels of accuracy, robustness and cybersecurity” for high-risk AI systems, including resilience against AI-specific attacks like data poisoning and model evasion. This is security engineering, not policy writing.
The shift from documentation to evidence
Traditional compliance follows a familiar pattern: write policies, conduct annual audits, produce documentation. The AI Act breaks this model. Article 74 gives market surveillance authorities power to demand:
- Access to training, validation, and testing datasets: not summaries, actual data
- Source code: may be requested on a reasoned, necessity-based basis after other reasonable means prove insufficient; access scope is role- and case-specific
- Logs demonstrating traceability: per Article 12, throughout the system lifecycle
- Operational records and test evidence, with effectiveness and coverage assessed separately
This creates a need for reviewable operational records alongside testing, technical documentation, and other evidence. Signed records can support integrity and attribution for covered fields; they do not by themselves prove that a control worked or that every relevant action was captured.
Key CISO responsibilities under the Act
The regulation assigns several technical domains squarely within CISO purview. Understanding these responsibilities is essential for resource planning and stakeholder communication.
1. Technical control implementation (Articles 9, 14, 15)
Article 9: Risk Management System
CISOs must implement continuous, iterative risk management including:
- → Identification and analysis of known and foreseeable security risks
- → Estimation and evaluation of risks that may emerge during deployment
- → Evaluation of risks based on post-market monitoring data
- → Adoption and documentation of suitable risk mitigation measures
Article 14: Human Oversight
Security implications of human oversight requirements:
- → Access controls ensuring authorized personnel can override AI decisions
- → Audit trails of human interventions and override decisions
- → Authentication mechanisms for human oversight functions
- → Secure channels for escalation and intervention
Article 15: Cybersecurity Requirements
AI-specific security controls beyond traditional IT security:
- → Data poisoning protection: Integrity verification for training data pipelines
- → Model evasion defense: Robustness testing against adversarial inputs
- → Model extraction prevention: API rate limiting and query monitoring
- → Model weight protection: Encryption and access controls for model files
2. Logging and audit-trail infrastructure (Article 12)
Article 12 establishes system-specific automatic event-logging requirements that directly affect CISO infrastructure decisions.
Article 12 baseline and implementation choices
- • Automatic logging capabilities ensuring traceability throughout the AI system lifecycle
- • Logging level appropriate to intended purpose: more critical systems require more granular logging
- • Remote biometric identification only: Article 12(3) specifies the period of use, reference database, input data leading to a match, and persons involved in verification
- • Integrity and access: Protecting logs against unauthorized alteration and access is prudent security practice; Article 12 does not generally prescribe cryptography or tamper-evident storage
- • Retention: Providers and deployers have separate obligations; Article 26 generally requires deployers to retain logs under their control for at least six months unless other law provides otherwise
3. Security testing requirements
The AI Act mandates testing that goes beyond traditional penetration testing. CISOs must establish programs covering:
- Adversarial testing: Systematic evaluation of model behavior under attack scenarios
- Robustness testing: Verification that systems perform correctly with noisy, incomplete, or edge-case inputs
- Red team exercises: For systemic-risk AI systems, formal adversarial evaluation per Article 55
- Bias and fairness testing: Security implications of discriminatory outputs
4. Incident-response obligations
Article 73 creates new incident response requirements specific to AI systems. CISOs must integrate these with existing security incident management:
Serious Incident Reporting (15-Day Deadline)
A “serious incident” under Article 73 includes any incident leading to:
- • Death or serious damage to health
- • Serious and irreversible disruption of critical infrastructure
- • Infringement of Union-law obligations intended to protect fundamental rights
- • Serious damage to property or the environment
CISOs should establish classification criteria and reporting procedures ahead of the applicable high-risk date: 2 December 2027 for relevant Annex III duties or 2 August 2028 for relevant Annex I product-embedded duties.
Questions CISOs should be asking
Before regulators ask these questions, CISOs should be asking them internally. Use this framework to assess your organization’s AI compliance readiness:
AI System Inventory
- ? Do we have a complete inventory of all AI systems in production?
- ? Which systems fall under high-risk classification per Annex III?
- ? Are there shadow AI deployments outside IT governance?
Logging & Evidence
- ? Can we produce logs demonstrating AI system behavior on demand?
- ? Are our logs tamper-evident and protected against unauthorized access?
- ? Do we log human oversight interventions and override decisions?
Security Controls
- ? Have we tested our AI systems against adversarial attacks?
- ? Do we have controls preventing data poisoning in training pipelines?
- ? Are model weights and training data protected with appropriate access controls?
Incident Response
- ? Do we have AI-specific incident classification criteria?
- ? Can we meet Article 73’s immediate-reporting trigger and the applicable 2-, 10-, or 15-day outer limit?
- ? Have we established communication channels with national competent authorities?
Red flags indicating compliance gaps
These warning signs suggest an organization may not be ready for the applicable 2027 or 2028 high-risk date under the AI Omnibus:
No AI system inventory exists
If you don’t know what AI systems you have, you can’t classify them or implement controls. Shadow AI is a critical blind spot.
Logging is application-level only
Generic HTTP logs may be insufficient, but Article 12 does not generally require every input, output and decision trace. Define events around intended purpose, traceability, risk identification and post-market monitoring.
Security testing excludes AI-specific threats
Penetration tests that don’t cover adversarial ML, data poisoning, or model extraction leave critical gaps.
AI governance is Legal’s responsibility alone
The AI Act requires technical controls that Legal can’t implement. Without CISO involvement, compliance is policy-only.
No budget allocated for AI compliance infrastructure
Article 12 logging infrastructure and Article 15 security controls require investment. Unfunded mandates don’t get implemented.
Personal-liability considerations for CISOs
While the EU AI Act primarily targets organizations with fines up to €35 million or 7% of global revenue, CISOs face personal liability exposure through several mechanisms:
Director and officer liability
Potential individual exposure depends on national corporate, criminal, professional, and sector law. AI Act organizational duties may be relevant evidence in a particular case, but they do not create one universal personal standard of care.
Criminal liability
The AI Act requires Member States to set effective, proportionate and dissuasive penalty rules, while national sanctions and procedures can differ. CISOs should confirm the current implementation and sector-specific consequences in each relevant jurisdiction.
Professional negligence
Failure to implement reasonable security controls for AI systems could expose CISOs to professional negligence claims, particularly if they were aware of risks and failed to act.
CISO Liability Mitigation Strategies
- → Document recommendations: Create written records of security recommendations, especially when budget or timeline constraints prevent implementation
- → Ensure board reporting: Regular reports on AI risk posture and compliance status create evidence of executive awareness
- → Review D&O insurance: Confirm coverage includes AI-related regulatory penalties and doesn’t exclude “regulatory compliance failures”
- → Establish governance structure: Formal AI governance committee with documented decision authority
Working with other stakeholders
EU AI Act compliance requires coordination across multiple functions. CISOs must establish effective working relationships with:
General Counsel (GC)
- • AI system classification and risk determination
- • Contract requirements for AI vendors
- • Incident reporting protocols and legal privilege
- • Authority information request responses
Chief Compliance Officer (CCO)
- • Quality management system integration
- • Conformity assessment preparation
- • Post-market monitoring coordination
- • Regulatory relationship management
Chief Technology Officer (CTO)
- • Technical documentation requirements
- • Logging infrastructure implementation
- • AI system architecture and security design
- • Adversarial testing program development
Board of Directors
- • AI risk appetite and tolerance definitions
- • Compliance investment authorization
- • Quarterly compliance status reporting
- • Material risk escalation decisions
Board and executive reporting requirements
CISOs should establish regular AI compliance reporting to the board. Recommended metrics and reporting elements:
Quarterly Board Report: AI Compliance Status
1. AI System Inventory Status
Total systems, classification by risk level, new systems added, systems retired
2. High-Risk System Compliance Progress
Percentage meeting Article 9-15 requirements, gap closure timeline, conformity assessment status
3. Technical Control Metrics
Logging coverage percentage, security testing completion, human oversight audit results
4. Incident Summary
AI-related incidents, near-misses, serious incident reports filed (if any)
5. Regulatory Engagement
Authority requests received, inspections, guidance documents reviewed
6. Material Risks and Recommendations
Identified compliance gaps, resource requirements, timeline risks
Implementation checklist for CISOs
Use this checklist to track your organization’s progress toward EU AI Act compliance:
EU AI Act Technical Requirements
Phase 1: Discovery
- Complete AI system inventory across all business units
- Classify systems per Annex III high-risk categories
- Identify shadow AI and unsanctioned deployments
- Assess current logging capabilities against Article 12
- Document existing security controls for AI systems
Phase 2: Infrastructure
- Implement logging infrastructure designed around the applicable Article 12 events
- Apply appropriate access, integrity and retention controls; consider signed records as an additional integrity measure
- Establish human oversight audit trail mechanisms
- Implement AI-specific security controls per Article 15
- Deploy training data integrity verification
Phase 3: Testing & Validation
- Establish adversarial testing program
- Conduct robustness testing for high-risk systems
- Validate logging completeness and accuracy
- Test incident response procedures
- Document testing results per Annex IV
Phase 4: Governance
- Establish AI incident classification criteria
- Create serious incident reporting procedures
- Implement board reporting framework
- Establish authority communication channels
- Document CISO recommendations and board responses
Timeline note: This 8-month timeline assumes dedicated resources and parallel workstreams. Work backward from 2 December 2027 for relevant Annex III duties or 2 August 2028 for relevant Annex I product-embedded duties.
Where GLACIS can support a CISO’s evidence work
For selected, configured workflow events, GLACIS can add signed operational records to a broader EU AI Act evidence program. Those records do not establish compliance, complete coverage, or control effectiveness:
Article 12 Logging Infrastructure
Logging designed around the events needed for traceability and risk monitoring for the particular high-risk system. Article 12 does not generally prescribe cryptography or require every input, output, and decision trace; signed records can support integrity for selected fields.
Continuous Control Attestation
Signed records of what a configured control reported on an in-scope path. Those records can support review, but separate testing is needed to establish control effectiveness and event coverage.
Reviewer-oriented reporting
Informational mappings and scoped metrics can help teams organize a board or reviewer discussion. They are not a compliance dashboard, legal determination, audit opinion, or substitute for the required technical and governance evidence.
Runtime coverage
Bounded operational evidence from a named workflow, assembled alongside the organization’s technical documentation and risk material. It supports review; it is not a compliance determination.
Frequently asked questions
What are the CISO’s specific responsibilities under the EU AI Act?
The AI Act assigns obligations to regulated organizations and defined operators, not a universal CISO job description. A CISO may own or support cybersecurity, logging, testing, and incident-response work, while legal, product, risk, and clinical teams retain other duties. Responsibility should be mapped to the organization’s role and the provisions that apply.
What logging requirements does Article 12 impose?
Article 12 requires high-risk systems to support automatic event logging over their lifetime to a degree appropriate to the system’s intended purpose, including events relevant to risk identification and post-market monitoring. Its more specific minimum fields apply to biometric-identification systems covered by Article 12(3). The Regulation does not generally prescribe cryptography, tamper-evident storage, or a complete input/output decision trace.
Can CISOs face personal liability under the EU AI Act?
The AI Act’s administrative penalties apply to regulated entities. Whether an individual CISO could face separate civil, employment, director, criminal, or insurance consequences depends on national law, role, conduct, and policy terms; the Act does not create a uniform personal-liability rule for CISOs. Obtain jurisdiction-specific advice.
How does the EU AI Act define cybersecurity requirements?
Article 15 requires high-risk AI systems to achieve appropriate levels of cybersecurity, including resilience against attempts to alter use, behavior, or performance through exploitation of vulnerabilities. This specifically includes technical solutions to address AI-specific vulnerabilities such as data poisoning, model evasion, adversarial attacks, and model extraction. Systems must also protect against unauthorized access to training data and model weights.
What is the serious incident reporting deadline?
Article 73 sets serious-incident reporting duties and timing that depend on the incident and system. CISOs should establish incident-classification procedures ahead of the applicable high-risk date (2 December 2027 for relevant Annex III duties, or 2 August 2028 for relevant Annex I product-embedded duties) and maintain communication channels with competent authorities.
How should CISOs coordinate with General Counsel?
CISOs should work with General Counsel on AI system classification and risk determination, contract requirements for AI vendors and deployers, incident reporting protocols and legal privilege considerations, documentation standards for regulatory defensibility, and coordinated responses to authority information requests under Article 74. Establish regular touchpoints and joint governance structures for effective collaboration.