Anthropic
ML/Research Engineer, Safeguards
$350K–$500K
San Francisco, CA +2 · Full-time · Posted Oct 9, 2025
- Location
- San Francisco, CA
New York City, NY
San Francisco, California, United States - Experience
- 4+ years in ML engineering, research engineering, or applied research; the posting also says required years will correlate with internal job-level requirements.
- Education
- Bachelor’s degree or an equivalent combination of education, training, and/or experience; relevant field of study demonstrated through coursework, training, or professional experience.
- Visa sponsorship
- Visa sponsorship is available, though not for every role or candidate; the company says it will make every reasonable effort to obtain a visa if it makes an offer.
About the job
Build and research systems that detect and mitigate misuse of AI, protect user wellbeing, and improve the safety of agentic products.
What you'll do
- Develop classifiers, synthetic data pipelines, and evaluation methods to detect misuse and anomalous behavior.
- Monitor harms across multiple exchanges and build methods to aggregate and analyze signals across contexts.
- Evaluate agentic-product safety and conduct research on mitigations, automated red-teaming, and adversarial robustness.
What they ask for
- 4+ years of experience in ML engineering, research engineering, or applied research.
- Proficiency in Python and experience building ML systems.
- Comfort working across the research-to-deployment pipeline.
- Strong communication skills, including explaining complex technical concepts to non-technical stakeholders.
Nice to have
- Experience with language modeling and transformers, classifiers, anomaly detection, behavioral ML, adversarial machine learning, red-teaming, interpretability, probes, reinforcement learning, or large-scale ML systems.
Pay
$350,000 — $500,000 USD annual base salary
Tools
Python, Language modeling, transformers, Reinforcement learning
AI focus
Develop safeguards to detect and mitigate AI misuse, monitor multi-exchange harms, test and mitigate agentic risks including prompt injection, and research automated red-teaming and adversarial robustness.
Model-written summary · the original posting is authoritative