Research Engineer, Safety & Alignment

Cognition, San Francisco

What you will do

Build agentic safety evals: Design and run evaluations for the failure modes that matter for autonomous agents: unsafe actions, prompt injection, data exfiltration, sandbox escape, reward hacking, and misuse. Make them fast enough to run on every model and product release.
Red-team Devin: Attack the agent and its harness systematically. Find the failures before customers do and turn them into fixes and regression tests.
Develop alignment methods: Work with post-training on reward modeling, preference data, constitutional approaches, and other techniques that make the model safer without making it worse at the job.
Shape the agent harness: Partner with the agent team on permissions, oversight, and escalation: when Devin should act, when it should ask, and how it should explain itself.
Publish and share: Contribute to Cognition's external safety work through papers, evals, and open methods where it makes sense.

Cognition, San Francisco

Build agentic safety evals: Design and run evaluations for the failure modes that matter for autonomous agents: unsafe actions, prompt injection, data exfiltration, sandbox escape, reward hacking, and misuse. Make them fast enough to run on every model and product release.
Red-team Devin: Attack the agent and its harness systematically. Find the failures before customers do and turn them into fixes and regression tests.
Develop alignment methods: Work with post-training on reward modeling, preference data, constitutional approaches, and other techniques that make the model safer without making it worse at the job.
Shape the agent harness: Partner with the agent team on permissions, oversight, and escalation: when Devin should act, when it should ask, and how it should explain itself.
Publish and share: Contribute to Cognition's external safety work through papers, evals, and open methods where it makes sense.