Junior AI Safety Engineer Resume Example

Best Practices for Junior AI Safety Engineer Resume

1. Open every bullet with a reproducible eval artifact. Replace 'tested AI for safety' with '47 jailbreak scenarios across 6 harm categories using HarmBench and PAIR templates'. The reproducibility is the whole point at junior level.
2. Quantify ASR, refusal recall, and false-positive rate. Even at junior, anchor every bullet with a number: ASR delta on a named harm class, refusal precision-recall on a sized prompt set, false-positive rate on a benign holdout. Numbers separate eval engineers from prompt taggers.
3. Name the harness, the model, and the harm class. Inspect AI on a Llama Guard 2 stack on the cybercrime harm class is the shape. Vague 'AI safety testing' phrasing reads as content moderation, not eval engineering.
4. Show the handoff. Trust and Safety reviewer, alignment-applied team, safety eval suite owner. Junior AI safety that does not feed signal to the model owner reads like an academic project.
5. Anchor to one harm taxonomy slot. Pick one harm class (cybercrime, CBRN, self-harm, persuasion) and run two bullets in it to show ownership of a slot, not random eval gigs.

Common Resume Mistakes for Junior AI Safety Engineer

1. Listing 'AI safety testing' without harm class, harness, or metric

Why it hurts: Recruiters at Anthropic, OpenAI, and DeepMind treat 'tested AI for safety' as noise. Without a named harm class, harness, and metric, the bullet is indistinguishable from content-moderation work.

How to fix: Replace 'tested AI for safety' with '47 jailbreak scenarios across 6 harm categories using HarmBench and PAIR, lifted ASR by 16 points'. Harness, harm class, count, delta. Four anchors, one bullet.

1. Confusing AI safety with cybersecurity or content moderation

Why it hurts: Junior resumes that lean on 'cybersecurity', 'compliance', or 'content moderation' framing get filtered into the wrong pile. AI safety hiring panels look for jailbreak/refusal/harm vocabulary, not CVE or trust-and-safety-ticket vocabulary.

How to fix: Rewrite the security or moderation bullets in eval-engineering terms. 'Triaged 800 abuse reports' becomes 'authored 32 reproducible refusal-recall test cases that surfaced a 6-point gap on the self-harm class'.

1. No reference to a real eval harness or guardrail

Why it hurts: Without Inspect AI, Eleuther LM-eval, simple-evals, Llama Guard 2, NeMo Guardrails, or Lakera Guard in the bullets, the work is invisible to senior eval engineers reviewing the resume.

How to fix: Pick one harness and one guardrail and place each inside an artifact. 'Implemented Eleuther LM-eval-harness wrapper for Llama Guard 2 on a 900-prompt dual-use eval set' is the form.

Quick Resume Tips for Junior AI Safety Engineer

1. Open with harness plus harm class plus delta. Inspect AI on cybercrime ASR is a one-line proof of competence.
2. Use the with-whom format. 'Co-authored a refusal rubric with the Trust and Safety reviewer' lands harder than 'helped on safety'.
3. Pair every tool with a release-gate outcome. HarmBench plus 'fed into pre-deployment red-team' is the shape.
4. Show one cross-team handoff per role. Trust and Safety reviewer, alignment-applied team, safety eval suite owner.
5. Keep one project on the resume that you can whiteboard end-to-end. Pick a HarmBench scenario pack or a Llama Guard 2 wrapper you can talk about for 25 minutes.