• Build adversarial environments for agents: complex, uncertain settings that sit on the boundary of agent capability and alignment, where failure is informative rather than trivial.
• Build realistic multi-agent environments and instrument them so emergent breakdowns are observable — failures that arise from the agents themselves, not ones scripted from the outside.
• Run experiments end to end, against external APIs and our own models, orchestrating many agents in parallel.
• Catalogue concrete agent failure modes and build the tooling to surface them at scale.
• Turn findings into internal models of agent behaviour and into public writeups.
• You’ll fit right in if you:
• Have built at least one non-trivial agent environment or automated research pipeline that ran end to end (single- or multi-agent), and can talk through what broke and why.
• Strong software and AI engineering. Can independently orchestrate many agents and containers in parallel without that orchestration being the bottleneck.
• A track record of empirical research in agents, red-teaming, or post-training where you defined the question, ran it, and drew a defensible conclusion.
• A fast empirical iterator who is comfortable defining the question when there's no playbook: can take a fuzzy concern ("do these agents collude under pressure?") and turn it into a concrete, falsifiable experiment.
• An AI power-user — fluent with frontier models and coding agents in your daily work.
• Published research at A* venues on automated red-teaming, agentic environments, or post-training.
• Experience building monitoring for model failures and anomalous behaviour.
• Experience reproducing public benchmark results and finding where the original methodology is fragile or misleading.
• An MSc or PhD in machine learning, computer science, cognitive science, computational neuroscience, physics, or a related quantitative field.
• AI safety fellowship (MATS, ASTRA, Anthropic Fellows, etc.), or a comparable self-directed research record.