13. AI Safety Engineer

Evidence label: Role family. Safety work appears under Software Engineer, Research Engineer, ML/Research Engineer, red-team, safeguards, and alignment titles. Evaluation is separated in Section 20 because it has a different production boundary.

What they do

Identify, measure, and reduce safety-relevant model or system behavior. The role can sit close to research, safeguards engineering, alignment, policy enforcement, or adversarial testing. It differs from AI Security Engineering: safety asks whether the system causes unacceptable behavior or harm; security asks how an attacker or compromised component can violate confidentiality, integrity, authorization, or availability.

Responsibilities

• Red-team AI systems to find failure modes, jailbreaks, and harmful outputs
• Build or validate safeguards, classifiers, output controls, and escalation paths
• Design human-in-the-loop checkpoints for high-risk decisions
• Define threat models and evaluations for misuse, alignment, or domain-specific harm
• Monitor deployed systems for safety regressions and emerging abuse patterns
• Translate findings into model, product, policy, or deployment changes

Required skills

Experimental design, statistics, Python, adversarial testing, model-behavior analysis, and clear risk reporting. Research-facing roles may require deeper ML or publication experience; production-facing safeguards roles may require stronger backend and incident-response skills.

Where to find these roles

Anthropic's current careers page includes safeguards, alignment, frontier red-team, and safety-research engineering titles. OpenAI and Google DeepMind also divide safety work across research and software-engineering families. These examples establish the work and title family, not a market-wide hiring rate.