AI Safety Research Engineer

Location: San Francisco, United States; work arrangement: on-site

Company

Windsurf builds Devin, an AI software engineer that writes and runs code, uses tools, and takes actions in real customer systems.

What you will do

• Build and run agentic safety evaluations for unsafe actions, prompt injection, data exfiltration, sandbox escape, reward hacking, and misuse.
• Red-team Devin and its harness, identify failures before customer impact, and turn findings into fixes and regression tests.
• Develop alignment methods involving reward modeling, preference data, constitutional approaches, and related techniques.
• Partner with the agent team on permissions, oversight, escalation, and decisions about when Devin should act or ask for help.
• Contribute to external safety work through papers, evaluations, and open methods where appropriate.

Requirements

• Hands-on experience in safety or alignment research, including evaluations, red-teaming, alignment, or interpretability.
• Understanding of agentic system failure modes such as tool misuse, specification gaming, long-horizon drift, and adversarial inputs.
• Strong engineering fundamentals with Python and PyTorch or JAX.
• Ability to build evaluation infrastructure, run experiments at scale, and read training and inference code.
• Strong empirical rigor and comfort with adversarial thinking.
• Experience at a frontier AI lab, applied AI company, or developer tools company; PhD in a relevant discipline or equivalent industry research experience.

Culture & Benefits

• Work as a founding member of the safety team, reporting to the Head of Safety.
• Safety research is integrated into a production AI agent operating at scale.
• Work alongside experienced researchers, engineers, founders, and competitive programmers from leading AI and technology companies.
• Equal opportunity employer with reasonable accommodations available throughout the hiring process.