“AI safety engineer” is a 2026 specialty that did not exist as a job title in 2020. The work spans red-teaming, guardrail engineering, evaluation, and the broader alignment program at frontier labs. Interviews probe a different set of skills than ML or applied-AI engineering. This guide is for the candidate considering the move and the EM trying to calibrate offers.
The companies hiring
- Frontier AI labs: Anthropic, OpenAI, Google DeepMind, Meta — significant safety teams
- Specialty safety / alignment companies: METR, Apollo Research, Arcadia, Aligned AI
- AI-shipping product companies: most have at least one safety hire by 2026
- Government and policy bodies: AISI (UK), AI Safety Institutes (US, others)
What the role actually does
- Red-teaming — adversarial testing of models for harmful behaviors
- Guardrail engineering — runtime classifiers, system-level filters, refusal behavior
- Evaluation — building benchmarks for safety properties
- Constitutional / RLHF safety training — shaping behavior through alignment techniques
- Interpretability — understanding why models do what they do
- Policy and incident response — when something does go wrong
Skills to know going in
- Strong Python and ML literacy — same as ML engineer
- Adversarial mindset — security/red-team background helps
- Statistical fluency for evaluation
- Familiarity with the alignment literature — RLHF, Constitutional AI, debate, etc.
- Comfort with ambiguity — many problems have no settled answer
Common interview rounds
Red-team scenario design
- “Design a red-team test suite for a coding assistant”
- “How would you uncover deceptive behavior in a model?”
- “What categories of attack would you prioritize for a customer service agent?”
Strong answers include attack-surface mapping, threat modeling, and the difference between automated and human red-teaming.
Guardrail design
- “Design a runtime guardrail for a customer-facing LLM that prevents PII leakage”
- “How would you build a refusal-classifier for sensitive medical advice?”
- “Walk me through latency vs safety tradeoffs in a streaming assistant”
Evaluation design
- “How would you evaluate whether a model has improved at refusing harmful requests?”
- “What is wrong with this benchmark? [shows a specific eval]”
- “Design an eval for jailbreak robustness”
Coding
- Often Python with statistics or NLP flavor
- Implement a small adversarial prompt generator
- Build a classifier for harmful content from labeled data
- Less LeetCode-heavy than ML or general engineering interviews
The “what would you do” round
Common at safety-mission companies (Anthropic especially):
- “You discover the model can be jailbroken into giving uplift on bioweapon synthesis. What do you do?”
- “A customer asks for a custom safety classifier that you suspect they will misuse. How do you respond?”
- “You find an interpretability result that suggests deceptive alignment. How do you act?”
These probe values and judgment more than skill.
The mission-fit interview
Anthropic, METR, and similar companies probe heavily for:
- Genuine concern about AI risks
- Calibrated views (not “AI is doom” or “AI is fine”)
- Track record of pursuing the work over years, not opportunism
- Willingness to tolerate ambiguity and slow feedback loops
Performative alignment is detected easily.
Compensation
- Anthropic / OpenAI safety roles: $300K–$700K total at senior; $700K–$1.5M+ at staff/principal
- Specialty companies (METR, Apollo): mid-tier cash; equity smaller; mission-driven compensation profile
- Government / AISI: lower base; but senior policy and impact
- Product-company safety roles: senior-IC band
How to break in
- Read the alignment literature: Anthropic’s research blog, the AI Alignment Forum, key papers (RLHF, Constitutional AI, Sleeper Agents, Activation Patching)
- Build a public red-team writeup or eval — find a specific failure mode and document it
- Contribute to OSS safety tooling (METR’s evals, lm-evaluation-harness, jailbreak benchmarks)
- Apply with a written portfolio, not just a resume
The career math
- Pros: meaningful work, well-funded teams, frontier-AI access, strong colleagues
- Cons: slow research cycles compared to product engineering, ambiguity, smaller external audience for your work
- Risks: field could mature in unexpected directions; specialty could narrow or broaden
What separates senior safety engineers from junior
Junior safety engineers run prescribed red-team test suites. Senior safety engineers design new evaluations, identify novel attack surfaces, and contribute to the methodology of the field. Staff safety engineers shape the safety roadmap of a frontier lab and write papers that influence industry-wide practice.
Frequently Asked Questions
Do I need an alignment PhD?
No. Strong ML / engineering background plus public safety work is sufficient for most roles. PhD helps for research-scientist-level safety roles.
Is this work durable?
Yes. The need for safety engineering grows with model capability. The specialty is unlikely to shrink for at least the next several years.
What about red-teaming consultancies vs in-house?
Both exist. In-house tends to pay more cash; consultancies offer breadth across many models. Many engineers do both at different career stages.