Head of Safety

About the role

Cognition is an applied AI lab building end-to-end software agents, including Devin, an autonomous agent that writes and runs code, uses tools and acts in real customer systems. The Head of Safety builds Cognition's safety function from the ground up, working directly with the researchers training the models and the engineers building the agent harness. The role is based in San Francisco.

Key responsibilities

• Set the safety strategy for Devin and its underlying models and own the outcomes
• Design evaluations and red-teaming for agentic risks such as unsafe actions, prompt injection, data exfiltration, sandbox escape, reward hacking and misuse, and integrate them into every model and product release
• Partner with pre-training, post-training and agent teams so safety is built into how models are trained and how the agent plans, acts and asks for help
• Define deployment policy: what the agent is and is not allowed to do, how permissions and oversight work, and how to handle gray areas in a way that satisfies enterprise customers
• Hire and lead safety researchers and engineers, and represent the safety work with customers, policymakers and the research community

Required profile

• Hands-on safety, alignment, evaluation or red-teaming experience at a frontier AI lab, with shipped work affecting real models or products
• Understanding of how agents fail in practice (tool misuse, specification gaming, long-horizon drift, adversarial inputs) and ideas on how to measure and mitigate it
• A track record of published or widely used safety research, evaluations or methods; an advanced degree in Computer Science, Machine Learning or a related field is a plus
• Experience starting or scaling a safety function, preferring shipping mitigations to writing memos
• Ability to make clear calls on deployment risk with incomplete information and explain them to engineers, executives and customers

Required skills

• Comfortable in Python and in the training and inference stack
• Able to read code, run evaluations and discuss details with researchers