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, known for creating Devin, the first AI software engineer. The founding team includes world-class competitive programmers, former founders, and leaders from Scale AI, Palantir, Cursor, Waymo, Tesla, Google DeepMind, and other cutting-edge AI companies.
As a Research Engineer on the Safety & Alignment team, you will be a founding member reporting to the Head of Safety. This is a hands-on role where safety work runs in the loop of a real production agent, not in benchmarks. Devin writes and runs code, uses tools, takes actions in real customer systems, and operates for hours without human oversight—making safety a practical production concern, not just research.
Key Responsibilities:
- Build agentic safety evaluations: Design and run evaluations for failure modes critical to autonomous agents, including unsafe actions, prompt injection, data exfiltration, sandbox escape, reward hacking, and misuse. Make evaluations fast enough to run on every model and product release.
- Red-team Devin: Systematically attack the agent and its harness to find failures before customers encounter them, then convert findings into fixes and regression tests.
- Develop alignment methods: Work with post-training on reward modeling, preference data, constitutional approaches, and other techniques that improve safety without degrading job performance.
- Shape the agent harness: Partner with the agent team on permissions, oversight, and escalation mechanisms—determining when Devin should act autonomously, when it should ask for help, and how it should explain its reasoning.
- Publish and share: Contribute to Cognition's external safety work through papers, evaluations, and open methods where appropriate.
Requirements:
- Demonstrated safety or alignment research experience: Hands-on work on evaluations, red-teaming, alignment, or interpretability at a frontier AI lab or in published research. Evidence of building systems that changed how a model was trained or deployed.
- Agentic systems understanding: Knowledge of how agents fail in practice (tool misuse, specification gaming, long-horizon drift, adversarial inputs), with measured experience or concrete ideas for measurement.
- Strong engineering fundamentals: Proficiency in Python and PyTorch (or JAX). Ability to build evaluation infrastructure, run experiments at scale, and read training and inference code.
- Empirical rigor: Design clean experiments, report results honestly, and distinguish real improvements from noise.
- Comfort with adversarial thinking: Enjoy breaking systems and demonstrate skill at it.
- Relevant industry experience: Prior experience at a frontier AI lab, applied AI company, or developer tools company.
- Advanced degree: PhD in Computer Science, Machine Learning, Physics, Mathematics, or related discipline; or equivalent industry research experience.
