AI Safety Research Scientist: Adversarial Environments

White Circle; Paris; Hybrid; EUR 90,000 - 150,000; Full time

Benefits offered by this job

Relocation package

Hybrid work in Paris

Comprehensive medical insurance

Job summary

White Circle is seeking a research scientist to build autonomous, large-scale environments that push LLM agents to failure and study how they break. You will develop adversarial, multi-agent environments and run end-to-end experiments against our models and external APIs.

Ideal candidates have MSc/PhD in ML/CS or related fields, a track record in empirical research on agents or post-training, and fluency with frontier models. Relocation to Paris is supported; hybrid work from Paris is available.

Qualifications

• Experience building agent environments or end-to-end research pipelines.
• Ability to orchestrate many agents and containers in parallel.
• Track record in empirical research with agents, red-teaming or post-training work.
• Fluency with frontier models and coding agents in daily work.

Responsibilities

• Build adversarial environments for agents near the boundary of capability and alignment.
• Create realistic multi-agent environments and instrument them to observe emergent breakdowns.
• Run end-to-end experiments against external APIs and our models with multiple agents in parallel.
• Catalogue concrete agent failure modes and surface them at scale.
• Turn findings into internal models and public writeups.

Skills

Agent environments

Software engineering

Empirical research

Experimentation

Frontier models

Education

MSc/PhD in ML or CS

AI safety fellowship (optional)

Job description

White Circle is seeking a research scientist to build autonomous, large-scale environments that push LLM agents to failure and study how they break. You will develop adversarial, multi-agent environments and run end-to-end experiments against our models and external APIs.

Ideal candidates have MSc/PhD in ML/CS or related fields, a track record in empirical research on agents or post-training, and fluency with frontier models. Relocation to Paris is supported; hybrid work from Paris is available.