How we assess
Every applicant goes through 5 gates: application, an AI-assisted screen, a live technical assessment scored by a practice lead, a client-fit interview, and reference and credential checks. The questions below come from the live assessment. We score answers on reasoning and trade-offs, not on recalling definitions.
20 technical questions with model-answer hints
Questions 1–10 (role-specific)
Question 1: Direct versus indirect prompt injection: what is the difference?
Model-answer hint
malicious user input versus instructions hidden in retrieved documents, emails or tool results.
Question 2: How do you test a RAG app for data exfiltration?
Model-answer hint
access-control bypass, cross-tenant retrieval, markdown or image exfiltration channels, log review.
Question 3: Which guardrails stop an agent misusing a tool?
Model-answer hint
least-privilege tools, argument validation, allow-lists, human confirmation for risky actions.
Question 4: Why can a system prompt not protect secrets?
Model-answer hint
extraction attacks succeed; keep secrets out of prompts and enforce rules in code.
Question 5: How do you automate jailbreak testing without drowning in false positives?
Model-answer hint
curated attack suites, judges calibrated on human labels, severity thresholds.
Question 6: Which OWASP LLM Top 10 risks matter most for an internal copilot?
Model-answer hint
prompt injection, sensitive information disclosure, excessive agency, supply chain.
Question 7: How would you detect poisoning in a fine-tuning dataset?
Model-answer hint
provenance checks, anomaly detection, trigger-phrase tests, held-out behaviour evaluations.
Question 8: How do output filters fail, and how do you layer them?
Model-answer hint
encoding tricks and language switching; combine classifiers, rules and structural constraints.
Question 9: What should be logged for an AI incident investigation?
Model-answer hint
prompts, retrieved sources, tool calls, model version and user identity, with privacy controls.
Question 10: How do model upgrades affect security posture?
Model-answer hint
behaviour shifts can reopen fixed issues; rerun the adversarial regression suite.
Questions 11–20 (Cybersecurity core)
Question 11: Map a control you implemented to NCA ECC-2:2024 or ISO 27001.
Model-answer hint
control reference, implementation, evidence produced, how effectiveness was tested.
Question 12: Walk through a ransomware intrusion using MITRE ATT&CK.
Model-answer hint
initial access to impact, with a detection or control at each stage.
Question 13: How do you prioritise 5,000 open vulnerabilities?
Model-answer hint
known exploitation, asset criticality, exposure, compensating controls, SLA tiers.
Question 14: Which logs do you need to investigate a compromised cloud identity?
Model-answer hint
sign-in, token issuance, control-plane audit, resource access and network flow logs.
Question 15: Explain zero trust to a CFO in 2 minutes.
Model-answer hint
no trust by network location; verify identity, device and context per request; frame as business risk.
Question 16: How do you write a detection that is useful, not noisy?
Model-answer hint
clear hypothesis, right data source, baseline tuning, emulation testing, documented triage steps.
Question 17: What do you do in the first hour of a suspected breach?
Model-answer hint
preserve evidence, contain, follow the playbook, start the regulatory clock, control communications.
Question 18: How do you protect secrets across CI/CD, cloud and laptops?
Model-answer hint
central vault, short-lived credentials, OIDC federation, secret scanning, rotation.
Question 19: How does SAMA CSF differ from ISO 27001 in practice?
Model-answer hint
sector regulator and maturity scoring versus a certifiable management system; large control overlap.
Question 20: Describe a security fix you won despite pushback.
Model-answer hint
quantified risk, business framing, agreed compromise, verified closure.
Interviewer tip: Ask for a real example after every answer. Strong candidates describe what they built, what broke and what they changed.
Practical task
Using a demo RAG assistant with a mock email tool, find and document at least 3 vulnerabilities, including one indirect prompt injection, rate their impact, and propose a guardrail, a design fix and a CI test for each.
How to run it Share the brief 24 hours ahead or run it live. Ask the candidate to talk through trade-offs as they work; reasoning matters more than a finished artefact.
Red flags our practice leads watch for
- Believes model alignment alone makes an application safe.
- Tests only with public jailbreak prompts, never application-specific attacks.
- Cannot explain trust boundaries in RAG or agent architectures.
- Produces findings that cannot be reproduced.
- Universal flag: Vague answers that never reach a concrete system, number or decision the candidate personally owned.
The scorecard we use
Score each dimension 1–4 (1 = no evidence, 2 = partial, 3 = solid, 4 = exceptional):
| Dimension | Weight | What a 4 looks like |
|---|---|---|
| Technical depth, production judgement and security | 40% | Explains how and why, designs for failure and sensitive data, and knows each tool’s limits |
| Practical task and problem solving | 30% | Delivers a working solution with tests and documented decisions |
| Communication and documentation | 20% | Explains clearly to technical and non-technical audiences, in English or Arabic |
| References and regional context | 10% | References confirm the claimed scope; knows the frameworks and data-residency rules that apply |
Pass bar: weighted average of 3.0 out of 4 or higher, with no dimension below 2. The practice lead approves the final result.