- 01Article 15 (accuracy, robustness, cybersecurity) gets less attention than the prohibited-use list but has the most concrete engineering implications: explicit obligations against data poisoning, model poisoning, evasion, and confidentiality attacks.
- 02Required engineering work: adversarial testing for prompt injection / jailbreaks / tool-use abuse; data lineage and provenance to detect poisoning; signed models with runtime attestation; logging that supports post-incident reconstruction of model behavior.
- 03Risk management is required throughout the lifecycle, not just at deployment. That maps to staged evaluations (pre-train, post-train, pre-deploy, in-production) and a documented re-evaluation trigger list.
- 04Most obligations land on existing AppSec, vuln management, and SIEM functions — but model integrity, data lineage, and prompt-layer monitoring need new tooling. Treat AI security as a distinct CISO program domain, not a side activity for the data team.
Most coverage of the EU AI Act focuses on prohibited uses, transparency, and high-risk classifications. Article 15 — accuracy, robustness, and cybersecurity — has gotten less attention but has the most concrete engineering implications.
Article 15 cybersecurity obligations
- 01Resilience to errors, faults, inconsistencies — including adversarial inputs
- 02Resilience to attempts by unauthorized third parties to alter use, outputs, or performance
- 03Specific controls against data poisoning, model poisoning, model evasion, and confidentiality attacks
- 04Risk management throughout the AI system's lifecycle, not just at deployment
Concretely, what does this require?
Adversarial testing for prompt injection, jailbreaks, and tool-use abuse for generative systems. Data lineage and provenance tracking sufficient to detect poisoning. Model integrity verification (signed models, runtime attestation). Logging that supports post-incident reconstruction of model behavior.
Map to your existing program
Most of these obligations land cleanly on existing AppSec, vulnerability management, and SIEM functions — but the AI specifics (model integrity, data lineage, prompt-layer monitoring) require new tooling. Treat AI security as a distinct program domain reporting into the CISO, not as a side activity for the data science team.