Solve hard problems where AI meets cybersecurity.

What you'll work on

• Build the runtime classification stack that decides whether an agent's proposed action is authorised before it executes.
• Design action-boundary controls that allow, block, or redact sensitive data from flagged actions.
• Build evidence pipelines that record every AI action, verdict, and rationale.
• Compile natural-language company policies and regulation into executable runtime controls.
• Develop multimodal ingestion and detection across models, tools, and data types.
• Own systems end to end and ship them into real, high-consequence customer deployments.

What we're looking for

A technically exceptional builder who cares about getting AI safety and security right, and who grounds research in real-world impact.

• Typically 0-3 years of post-university experience building AI or machine learning systems.
• A strong foundation in computer science, machine learning, cybersecurity, or a related field.
• Evidence of exceptional building ability through production systems, research at a leading lab, startup work, or open source.
• A rigorous understanding of how AI systems fail and how adversaries can exploit them.
• The ability to move quickly, communicate clearly, and own complex technical problems independently.

What you'll work on

• Red-team production and pre-production agents across browser, code, enterprise, and multi-agent environments.
• Discover attacks involving prompt injection, goal manipulation, memory poisoning, tool abuse, identity, and data exfiltration.
• Develop realistic exploit chains across model, application, identity, tool, and infrastructure layers.
• Build automated attack agents, evaluation harnesses, and reusable security test suites.
• Turn research findings into detections, runtime evidence, policy controls, and product requirements.
• Assess customer deployments and communicate technically rigorous, actionable findings.

What we're looking for

An adversarial thinker with a builder's instinct who can move from an ambiguous attack hypothesis to a reproducible proof of concept.

• A strong foundation in security, computer science, machine learning, or a related field.
• Hands-on experience in offensive, application, cloud, product, or AI security.
• The ability to build security tooling and proof-of-concept exploits, ideally in Python and a systems language.
• Practical knowledge of modern agent architectures, including tool calling, retrieval, memory, orchestration, and permissions.
• Strong experimental discipline, clear communication, and ownership of technically ambiguous problems.