Software Engineer, Safeguards
Anthropic · San Francisco, CA | New York City, NY
About Anthropic Anthropic’s mission is to create reliable, interpretable, and steerable AI systems. We want AI to be safe and beneficial for our users and for society as a whole. Our team is a quickly growing group of committed researchers, engineers, policy experts, and business leaders working together to build beneficial AI systems. About the role We are looking for software engineers to help build safety and oversight mechanisms for our AI systems. As a software engineer on the Safeguards team, you will work to monitor models, prevent misuse, and ensure user well-being. This role will focus on building systems to detect unwanted model behaviors and prevent disallowed use of models. You will apply your technical skills to uphold our principles of safety, transparency, and oversight while enforcing our terms of service and acceptable use policies. Responsibilities: Develop monitoring systems to detect unwanted behaviors from our API partners and potentially take automated enforcement actions; surface these in internal dashboards to analysts for manual review Build abuse detection mechanisms and infrastructure Surface abuse patterns to our research teams to harden models at the training stage Build robust and reliable multi-layered defenses for real-time improvement of safety mechanisms that work at scale You may be a good fit if you: Bachelor’s degree in Computer Science, Software Engineering or comparable experience 5-10+ years of experience in a software engineering position, preferably with a focus on integrity, spam, fraud, or abuse detection and mitigation Proficiency in Python and Typescript Ability to work across the stack Strong communication skills and ability to explain complex technical concepts to non-technical stakeholders Strong candidates may also: Have experience building trust and safety detection mechanisms and intervention for AI/ML systems Have experience with prompt engineering, jailbreak attacks, and other adversarial inputs Have worked