Safety Research Scholar - Human Frontier Labs

About the Program The Human Frontier Collective (HFC) brings researchers and scientists together to tackle consequential problems in AI. In partnership with Scale Labs, we are building a small cohort of researchers to work on the empirical foundations of AI safety. That work includes finding where models fail, checking whether safety controls hold up under pressure in realistic settings, and studying how AI agents (mis-)behave over long-horizon tasks. Scholars take open research questions like these and turn them into work that improves how frontier AI systems are evaluated and safeguarded. Each scholar owns a substantial part of a safety research project. You'll work alongside a mentor from the research question through the experiments to deciding what to test next. Research Areas Mentors will guide Research Scholars on projects in selected AI safety research areas in select AI safety research areas, such as: • AI control and chain-of-thought monitoring: Developing techniques to oversee and constrain AIs, including detecting harmful intent or deception. • Alignment auditing and evaluations: Building scalable methods to test whether models have hidden goals, misaligned behaviors or unsafe tendencies. • Multi-agent safety: Studying the risks that emerge when multiple agents interact, such as collusion, coordinated failures, or unexpected collective behavior. • Recursive Self-Improvement (RSI) safety: Investigating how to keep AI systems safe and aligned as AIs increasingly contribute to their own development. • Science of Evaluations: examining when and how models recognize they're being tested, and how situational awareness affects the reliability/validity of evaluations. • Misuse risks and evaluations: Developing methods to assess how AI systems could be exploited by malicious actors, including evaluating their potential to enable harmful activities. Projects are matched to demonstrated expertise; we don’t expect any candidate to cover every area. Representative Research • ASPI : 728 task–attack scenarios testing how clarification-seeking changes prompt-injection vulnerability in LLM agents. • PropensityBench : simulating real-world pressure to choose between safe or harmful behavior • BioRiskEval : evaluating whether safeguards for open-weight bio-foundation models remain robust under fine-tuning and probing. • See more work from the Human Frontier Collective . What You'll Do Projects are mentor-guided. By default, you will be given a list of research proposals and match you with one based on your research interests and background. You’ll meet weekly with a Scale researcher who sets the direction and gives feedback, and you will drive the day-to-day work. • Own research end to end: Turn a high-level safety question into hypotheses, baselines, and controlled experiments, and produce work that others can inspect and build on. • Build the measurement: Implement and debug evaluation harnesses, agent environments, datasets, monitoring methods, or post-training experiments. • Exercise research judgment: Identify confounds, separate what a model can do from what it tends to do, make sense of negative results, and decide what to try next without step-by-step direction. • Make the work legible: Produce reproducible code, experiment logs, analysis, and precise technical writing. External outputs, where appropriate, remain subject to research, security, and communications review. Outputs also include research papers, which will be submitted to top ML conferences (NeurIPS, ICML, ICLR). Who We're Looking For You may be a good fit if you: • are motivated by reducing catastrophic risks from advanced AI systems; • are interested in technical safety research, such as alignment evaluations, AI control, monitoring, multi