About the role
About the Team
The Guardian team builds practical safety and oversight systems for increasingly capable AI agents. Our goal is to make it possible for agents to take on longer-running, higher-autonomy work while preserving meaningful safeguards, respecting user intent, and intervening when actions could cause serious or difficult-to-reverse harm.
Auto-review is an early example of this work. When an agent reaches a permission boundary, a separate model evaluates the proposed action in context and determines whether it should proceed. The team is now extending this foundation into more capable oversight systems that account for richer user context, support a broader range of agent experiences, and integrate across OpenAI products.
Guardian sits at the intersection of engineering, alignment research, product, infrastructure, and security. The team already has a strong engineering foundation in London and will continue to grow there as a concentrated, highly collaborative group.
About the Role
We are looking for a Software Engineer to join the Guardian team in London. You will design and build production systems that evaluate agent behavior, identify risk, and take appropriate action without creating unnecessary friction for users.
This is a hands-on engineering role with substantial technical depth and ownership. You will work across architecture, implementation, evaluation, and operations, taking projects from early exploration through production deployment. The work requires strong engineering fundamentals, sound product judgment, and comfort solving problems where both the requirements and the underlying technology are evolving.
You will work closely with alignment researchers to translate emerging research into dependable systems, and with product and engineering teams to make agent safety a foundational capability across OpenAI.
This role is based in our London office.
In this role, you will
- Design, build, and operate core components of Auto-review and the broader Guardian platform, balancing safety, product utility, latency, and operational reliability.
- Build production systems that evaluate agent actions in real time, detect potentially harmful behavior, and support safe recovery or escalation paths.
- Improve how oversight systems understand user intent, environmental context, permissions, and relevant personalization signals while protecting privacy and avoiding unnecessary friction.
- Partner closely with alignment researchers to turn prototypes, evaluations, and new monitoring techniques into scalable product capabilities.