Teams at Refractal

Lead offensive security research against AI agents and their infrastructure. Red-team browser, code, enterprise, and multi-agent systems; investigate prompt injection, tool abuse, memory poisoning, identity attacks, and data exfiltration. Build automated attack tooling and evaluation suites, convert findings into detections and controls, and lead customer security assessments. The role requires original security research, hands-on LLM application testing, strong Python and systems programming, and clear communication.

Conduct authorized red-team assessments of AI applications and agents, testing threats such as prompt injection, tool misuse, data exposure, and permission-boundary violations. Build reproducible testing tools and regression suites, analyze interactions among models, applications, identities, and permissions, document practical security findings, recommend mitigations, and retest fixes with engineering teams.

Own production machine-learning models that detect prompt injection, data exfiltration, policy violations, and other AI threats. Train and deploy low-latency classifiers, analyze telemetry for drift and false positives, build evaluations and benchmarks, and convert new attacks into training data. The role also translates policies and regulations into measurable controls while improving model quality and latency. Candidates need substantial applied ML or data science experience, strong Python and PyTorch skills, NLP and LLM expertise, evaluation statistics, and clear communication.