Interview dimensions are easier to defend when they are not your own invention. Each of the eight skills maps onto published standards, so a hiring manager can justify the stage to a CTO or a head of risk by pointing at the standard rather than at a framework.
The mapping is not a syllabus. No candidate should be asked to recite an OWASP entry, and someone who has never read any of these documents can still answer every question well. The standards describe the failures; the questions test whether someone has met them.
Skills to OWASP and NIST
| Skill | OWASP Top 10 for LLM Applications (2025) | NIST function |
|---|---|---|
| Non-Determinism Engineering | LLM09 Misinformation | Measure, Manage |
| Knowledge Governance | LLM08 Vector and Embedding Weaknesses; LLM09 Misinformation | Map |
| Autonomy Boundary Design | LLM06 Excessive Agency | Govern, Manage |
| Adversarial Input and Output Handling | LLM01 Prompt Injection; LLM05 Improper Output Handling; LLM07 System Prompt Leakage | Map, Manage |
| Data Boundary and Privacy Engineering | LLM02 Sensitive Information Disclosure; LLM08 Vector and Embedding Weaknesses | Map, Govern |
| Evaluation and Observability Design | LLM09 Misinformation (detection side) | Measure |
| Cost, Latency and Capacity Engineering | LLM10 Unbounded Consumption | Manage |
| Model Lifecycle and Supply Chain | LLM03 Supply Chain; LLM04 Data and Model Poisoning | Govern, Manage |
All ten OWASP entries are covered by at least one skill.
By NIST function
| Function | Skills |
|---|---|
| Govern | Autonomy Boundary Design, Data Boundary and Privacy Engineering, Model Lifecycle and Supply Chain |
| Map | Knowledge Governance, Adversarial Input and Output Handling, Data Boundary and Privacy Engineering |
| Measure | Non-Determinism Engineering, Evaluation and Observability Design |
| Manage | Non-Determinism Engineering, Autonomy Boundary Design, Adversarial Input and Output Handling, Cost, Latency and Capacity Engineering, Model Lifecycle and Supply Chain |
A candidate strong only in Measure is a good evaluator who has not shipped. A candidate strong only in Manage has operated something but may not know whether it is any good. The combination is what the cluster looks for.
Where the EU AI Act touches the engineering decision
GPAI obligations have applied since 2 August 2025. High-risk obligations, and the Commission's enforcement powers over general-purpose models, apply from 2 August 2026. Details are on the European Commission's AI Act page.
| Engineering decision | Why it matters | Skill |
|---|---|---|
| Classifying the feature | High-risk status follows the use case, not the model size. Employment, credit, education and essential services are the usual surprises | Data Boundary and Privacy Engineering |
| Record keeping | Reconstructing what the system did, months later, for someone who is not an engineer | Evaluation and Observability; Model Lifecycle |
| Human oversight | An oversight requirement is a design requirement. A human who cannot understand or override the output in time is not oversight | Autonomy Boundary Design |
| Transparency to users | Telling people they are interacting with a machine, and labelling generated content | Autonomy Boundary Design; Data Boundary |
| Documentation of models in use | Provider documentation becomes part of your own technical file | Model Lifecycle and Supply Chain |
The cluster does not test legal knowledge. It tests whether an engineer raises the question before launch, which is the behaviour the obligation actually depends on.
Sources
- OWASP Top 10 for Large Language Model Applications, 2025
- NIST AI 600-1, Generative AI Profile, July 2024
- Regulation (EU) 2024/1689, the AI Act
© Gabor Mayer. Licensed under Creative Commons Attribution 4.0 (CC BY 4.0). Free to share and adapt with attribution.