SlipstreamJobs tracks this role from the company's public career site. Apply directly on the employer's site.
Cognition is an applied AI lab building end-to-end software agents, creators of Devin, the first AI software engineer. The company is composed of world-class talent including competitive programmers, former founders, and leaders from frontier AI organizations including Scale AI, Palantir, Cursor, Waymo, Tesla, Google DeepMind, and others.
You will build Cognition's safety function from the ground up, owning the research agenda, evaluation and red-teaming program, deployment policies, and team building. Devin 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 human oversight. This makes it critical to get safety right, and one of the few places where safety work ships to millions of developers rather than remaining theoretical.
Key responsibilities:
- Own safety strategy end-to-end for Devin and its underlying models, with accountability for outcomes
- Design and build evaluation programs and red-teaming for agentic risks including unsafe actions, prompt injection, data exfiltration, sandbox escape, reward hacking, and misuse
- Integrate safety into every model and product release
- Partner with pre-training, post-training, and agent teams to embed safety into model training and how Devin plans, acts, and requests help
- Define deployment policies: what Devin can and cannot do, how permissions and oversight function, and how to handle gray areas in ways that satisfy enterprise customers
- Hire and lead safety researchers and engineers
- Represent Cognition's safety work with customers, policymakers, and the broader research community
Requirements:
- Demonstrated frontier lab safety or alignment experience: hands-on work on safety, alignment, evaluations, or red-teaming at a frontier AI lab with shipped safety work that affected real models or products
- Deep understanding of how agentic systems fail in practice (tool misuse, specification gaming, long-horizon drift, adversarial inputs) with ideas on measurement and mitigation
- Track record of published or widely used safety research, evals, or methods
- Advanced degree in Computer Science, Machine Learning, or related field (preferred)
- Technical fluency: comfortable in Python and the training/inference stack; able to read code, run evals, and engage researchers on technical details
- Builder mindset: preference for shipping mitigations over writing memos
- Sound judgment under uncertainty: ability to make clear deployment risk calls with incomplete information and explain them to engineers, executives, and customers
