About the role Join our new AGI safety product team and help transform AI control research into practical tools that directly reduce risks from AI. As an Research Scientist (Control), you’ll work closely with Marius (CEO & currently leads the monitoring efforts), other control researchers and product engineers. We are currently building Watcher, a monitoring tool for coding agents. Our monitoring research agenda attempts to translate compute into safety at scale. You will join a small team and will have significant ability to shape the team & tech, and have the ability to earn responsibility quickly. You will like this opportunity if you're passionate about using empirical research to make AI systems safer in practice. You enjoy the challenge of translating theoretical AI risks into concrete detection mechanisms. You thrive on rapid iteration and learning from data. You want your research to directly impact real-world AI safety. Responsibilities TLDR: you will design & implement control protocols (see e.g. [Greenblatt et al, 2023]) and test them on real-world production systems at scale. Research & Development Design and conduct experiments to test monitor effectiveness across different failure modes and agent behaviors Build and maintain evaluation frameworks to measure progress on monitoring capabilities Build and maintain high-quality datasets to train and test monitors on Iterate on monitoring approaches based on empirical results, balancing detection accuracy with computational efficiency Contribute newly discovered coding agent failure modes to Apollo's shared failure mode library (owned by the AI Security & Control Researcher/Engineer) as they surface during monitor design, evaluation, and iteration Stay current with research on AI safety, agent failures, and detection methodologies Stay current with research into coding security and safety vulnerabilities Fine-tuning & Monitor Development Fine-tune open-source models to create efficient monitors for high-volume production environments Design and build agentic monitoring systems that autonomously investigate logs to identify both known and novel failure modes Incorporate adversarial findings from red-teaming campaigns into monitor design and evaluation, closing the loop between attack and defense Requirements 2+ years of experience conducting empirical research with large language models or AI systems Strong experience with AI coding agents. For example, having extensively used and compared frontier coding agents, or having designed / developed coding agents Experience with LLM-as-a-judge setups or AI monitoring more broadly Experience designing and running experiments, analyzing results, and iterating based on empirical findings e.g. prompting, scaffolding, agent design, fine-tuning, or RL Strong Python programming skills Demonstrated ability to work independently on open-ended research problems Nice to have Familiarity with AI safety concepts, particularly agent-related risks Familiarity with computer security, e.g. security testing and secure system design Experience fine-tuning language models or working with smaller open-source models Previous work building developer tools or monitoring systems Publications or contributions to AI safety or ML research Experience with production log systems or production log analysis We want to emphasize that people who feel they don't fulfill all of these characteristics but think they would be a good fit for the position nonetheless are strongly encouraged to apply. Representative Projects Hierarchical monitoring for coding agent security: Design a multi-layer monitoring system for detecting security vulnerabilities introduced by coding agents. Start by cataloging common security failure modes (e.g., hardcoded credentials, SQL injection vulnerabilities, insecure API calls). Build specialized monitors for each category, then create a hierarchical system where fast, efficient first-pass monitors flag potentially problematic code for deeper investigation by more sophisticated monitors. Design a backtesting strategy for Watcher: Validate the system on synthetic test cases and real agent outputs, iterating to optimize the tradeoff between detection rates and false positives while maintaining low latency for most monitoring decisions. Fine-tune an open-source model to be a great monitor: Take an open-source model and fine-tune it on our large dataset of coding agent failures with high-quality ground truth labels. Test different fine-tuning techniques and measure generalization to our held-out data. Compare against prompted baselines on accuracy, cost and latency. The goal is to fine-tune models to lift the pareto frontier of monitors. Benefits This role offers market competitive salary, equity, and competitive benefits. Salary: San Francisco: $204,000 – $385,000; London: £136,000 – £258,000. We will be looking to meaningfully raise salaries soon. Flexible work hours and schedule Unlimited vacation Unlimited sick leave Up to 6 months of paid parental leave Comprehensive health, dental and vision insurance Retirement savings with competitive employer matching (e.g. 401(k) for US employees) Lunch, dinner, and snacks are provided for all employees on workdays Paid work trips, including staff retreats, business trips, and relevant conferences A yearly $1,000 (USD) professional development budget Relocation support and visa fees (if applicable) Logistics Time Allocation: Full-time Location: This is an in-person role working out of our London or San Francisco office. We offer flexible working hours and wfh arrangements. Visa sponsorship: We sponsor visas in both the UK and US. Sponsorship isn't guaranteed for every role or candidate, but if we make you an offer, we'll work with you to find the right visa route.
