Review the evidence signals before interviewing. Then use the anchored descriptions—not instinct alone—to choose the score that best matches each answer.
01
Evaluation factor
Technical depth
35% weight
Check depth on attack classes: PGD and AutoAttack evasion, data poisoning, membership inference, model extraction, plus prompt injection and jailbreak methods against LLM guardrails.
Evidence to listen for
Command of the specific attack surface, tooling, and controls the role covers
Understands how the underlying system works, not just how the tool reports on it
Can explain an attack or control chain end to end
Distinguishes what they found themselves from what a scanner flagged
Five-point scoring guide
1
Poor
Tool operator only; no understanding of the systems underneath.
2
Needs Improvement
Runs tooling but cannot explain findings or how the attack works.
3
Satisfactory
Solid working knowledge; depth thins outside familiar tooling.
4
Very Good
Strong command of the domain; explains attack and control chains clearly.
5
Excellent
Names specific threat models and epsilon budgets, distinguishes certified defences from gradient masking, and explains why adversarial training costs accuracy.
02
Evaluation factor
Real incidents and findings
30% weight
Probe concrete red-team engagements: models attacked, datasets poisoned, CVEs or MITRE ATLAS techniques mapped, attack success rates achieved, and findings filed against production classifiers.
Evidence to listen for
Brings specific incidents, findings, or audits they personally worked
States their own role rather than the team's
Describes what was actually at risk and what changed afterwards
Can talk about a finding that turned out to be wrong
Five-point scoring guide
1
Poor
No hands-on work; knowledge is entirely certification or coursework.
2
Needs Improvement
Limited exposure; cannot describe their contribution to an incident.
3
Satisfactory
Real casework with adequate detail; ownership sometimes vague.
4
Very Good
Specific incidents with clear personal scope and what changed after.
5
Excellent
Walks through a real engagement end to end with measured attack success rates, affected model versions, and the writeup that reached engineering.
03
Evaluation factor
Risk judgement
20% weight
Assess how they rank findings when a 90 percent white-box success rate collapses under realistic black-box, rate-limited, or physically constrained deployment conditions.
Evidence to listen for
Prioritises by actual exploitability and business impact, not raw severity scores
Can argue for accepting a risk as well as fixing it
Knows the difference between a finding and a problem
Does not cry wolf or wave things through
Five-point scoring guide
1
Poor
Treats every finding as critical, or waves real risk through.
2
Needs Improvement
Follows severity scores mechanically; no business context.
3
Satisfactory
Reasonable prioritisation; less confident arguing for risk acceptance.
4
Very Good
Prioritises by exploitability and impact; can justify accepting a risk.
5
Excellent
Separates lab-only attacks from deployable ones, argues threat realism from access assumptions, and declines to inflate severity on unreachable models.
04
Evaluation factor
Getting things fixed
15% weight
Look for evidence they drove fixes: input sanitisation, detection monitors, retraining pipelines, guardrail changes shipped with ML engineers, and regression tests preventing reintroduction.
Evidence to listen for
Writes findings engineers can act on rather than a wall of output
Has persuaded a team to fix something they did not want to fix
Explains risk to executives in business terms
Works with the org rather than policing it
Five-point scoring guide
1
Poor
Adversarial with engineering; findings never get fixed.
2
Needs Improvement
Reports are unactionable; no influence beyond raising tickets.
3
Satisfactory
Adequate reporting; relies on mandate rather than persuasion.
4
Very Good
Actionable findings and a real record of getting fixes shipped.
5
Excellent
Cites defences that shipped and stuck, including robustness evaluation harnesses added to CI and a measurable drop in attack success.
Put this rubric to work
Score every candidate against the same standard
Add these weighted factors to Hirevire and let AI evaluate recorded answers against your rubric.
