Article 15 is the EU AI Act requirement for accuracy, robustness, and cybersecurity in high-risk AI systems. It expects technical measures that prevent, detect, respond to, resolve, and control attacks such as data poisoning, model evasion, and confidentiality attacks, with evidence that those controls remain effective over time.
Expanded Definition
Article 15 is the EU AI Act’s baseline security and resilience requirement for high-risk AI systems. It is not limited to generic “secure by design” language. It specifically expects accuracy, robustness, and cybersecurity controls that can resist attacks against data, models, and outputs over time. That includes technical measures to prevent, detect, respond to, resolve, and control events such as data poisoning, model evasion, and confidentiality attacks.
In practice, Article 15 pushes organisations to treat AI security as an operational discipline with evidence, testing, and continuous monitoring rather than a one-time assurance claim. That aligns closely with the risk patterns described in the OWASP NHI Top 10 and with broader guidance from the CISA cyber threat advisories, both of which emphasise persistent attack paths and control validation. Definitions vary across vendors on how far Article 15 extends into organisational process, but no single standard governs this yet.
The most common misapplication is treating Article 15 as a documentation exercise, which occurs when teams produce risk statements without proving that controls still work after model updates, data drift, or attacker adaptation.
Examples and Use Cases
Implementing Article 15 rigorously often introduces ongoing validation overhead, requiring organisations to weigh model agility against the cost of continuous assurance, testing, and monitoring.
- A bank deploys anomaly detection and red-team testing to confirm a fraud model remains resilient after retraining and new data ingestion.
- A healthcare provider adds poisoning checks to training pipelines and monitors for drift that could degrade diagnostic output quality.
- An enterprise secures a customer support agent by limiting sensitive context exposure and reviewing prompts for confidentiality leakage risks, informed by the Ultimate Guide to NHIs — Key Challenges and Risks.
- A product team validates fallback behaviour and incident response playbooks after a model begins producing unsafe or manipulated outputs.
- Security teams compare offensive testing results against the MITRE ATLAS adversarial AI threat matrix to prioritise likely attack paths.
These use cases are especially important where AI systems depend on external tools, third-party data, or long-lived secrets. That reality is reflected in The 52 NHI breaches Report, which shows how identity and access weaknesses often amplify downstream compromise.
Why It Matters in NHI Security
Article 15 matters to NHI security because many high-risk AI systems rely on non-human identities, API keys, service accounts, and orchestration tokens to access data, tools, and downstream systems. If those identities are weakly governed, attackers can move from model abuse into broader system compromise. NHI security failures often become AI security failures, especially when secrets are embedded in pipelines, over-privileged, or not rotated.
NHI Management Group’s research in Ultimate Guide to NHIs shows that 96% of organisations store secrets outside secrets managers in vulnerable locations and 71% of NHIs are not rotated within recommended time frames. Those conditions directly undermine the durability Article 15 expects from cybersecurity controls. The same operational weakness appears in incident data from 52 NHI breaches Analysis, where identity compromise frequently accelerates wider breach impact.
Practitioners should view Article 15 as a requirement to prove control effectiveness over time, not merely to assert model quality at launch. Organisations typically encounter this requirement only after a model update, attack, or regulatory review exposes that their AI system’s supporting identities and secrets were never governed tightly enough, at which point Article 15 becomes operationally unavoidable to address.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
OWASP Agentic AI Top 10 and MITRE ATLAS address the attack surface, NIST AI RMF set the technical controls, and EU AI Act and EU Cyber Resilience Act define the regulatory obligations.
| Framework | Control / Reference | Relevance |
|---|---|---|
| EU AI Act | Article 15 | Directly defines the cybersecurity, robustness, and accuracy requirement for high-risk AI systems. |
| OWASP Agentic AI Top 10 | A2 | Agentic systems face prompt, tool, and output abuse patterns that map to Article 15 concerns. |
| MITRE ATLAS | Provides adversarial AI tactics and techniques relevant to poisoning, evasion, and extraction. | |
| NIST AI RMF | MAP, MEASURE, MANAGE | Frames AI risk management as a lifecycle practice with measurable controls and monitoring. |
| EU Cyber Resilience Act | Supports product security expectations for digitally embedded components and update resilience. |
Treat AI-enabled products as secure-by-design assets with patching, monitoring, and vulnerability handling.
