Role MissionDevin is one of the most capable autonomous agents in production. It writes and runs code, uses tools, takes actions in real customer systems, and operates for hours without a human in the loop. That makes it one of the most important places in the industry to get safety right, and one of the few where safety work ships to millions of developers rather than staying in a paper. You will build Cognition's safety function from the ground up: the research agenda, the evaluation and red-teaming program, the deployment policies, and the team. You will work directly with the researchers training our models and the engineers building the agent harness. This is a hands-on role for someone who has done serious safety or alignment work at a frontier lab and wants to apply it where agents actually operate.
What You'll Accomplish- Own safety end to end: Set the safety strategy for Devin and the models behind it, and own the outcomes.
- Build the evaluation program: Design evals and red-teaming for agentic risk: unsafe actions, prompt injection, data exfiltration, sandbox escape, reward hacking, and misuse. Make them part of every model and product release.
- Shape training and the agent harness: Partner with pre-training, post-training, and agent teams so safety is built into how models are trained and how Devin plans, acts, and asks for help.
- Set deployment policy: Define what Devin is and is not allowed to do, how permissions and oversight work, and how we handle the gray areas, in a way that holds up with enterprise customers.
- Build the team and the external voice: Hire and lead safety researchers and engineers. Represent Cognition's safety work with customers, policymakers, and the broader research community.
Exceptional Candidates Have Demonstrated- Frontier lab safety or alignment experience: Hands-on work on safety, alignment, evaluations, or red-teaming at a frontier AI lab. You have shipped safety work that affected real models or products.
- Agentic systems depth: You understand how agents fail in practice: tool misuse, specification gaming, long-horizon drift, adversarial inputs. You have ideas about how to measure and mitigate it.
- Research credibility: A track record of published or widely used safety research, evals, or methods. An advanced degree in Computer Science, Machine Learning, or a related field is a plus.
- Technical fluency: Comfortable in Python and in the training and inference stack. You can read the code, run the evals, and argue with researchers on the details.
- Builder, not reviewer: You have started or scaled a safety function and prefer shipping mitigations to writing memos about them.
- Judgment under uncertainty: You can make clear calls on deployment risk with incomplete information and explain them to engineers, executives, and customers.
