Research Scientist — AI Safety & Misalignment Evaluation

Visa Hunt; Paris; Hybride; EUR 90 000 - 140 000; Plein temps

Avantages offerts par ce poste

Relocation package

Hybrid Paris/London work option

Medical insurance (France)

Hardware and tools provided

Subscriptions for AI agents and IDEs

Team off-sites twice a year

Résumé du poste

White Circle is an AI Safety company focused on evaluating and improving safety, reliability, and optimization for AI systems. We design evals, benchmarks, environments, and tooling to study how AI systems fail and how to prevent unsafe behaviors in user-facing scenarios.

The research team conducts end-to-end studies—from vague questions to reproducible experiments—using white-box and black-box investigations to understand failures and biases in real-world interactions.

Qualifications

• Empirical research in agent behavior and alignment.
• Experience designing real-world experiments under varying conditions.
• Ability to turn vague behavioural questions into concrete experiments and fast iteration.
• Familiarity with frontier models, AI safety concepts, and interpretability tooling.

Responsabilités

• Own research projects end to end—from unclear concerns to falsifiable experiments, baselines, and defendable results.
• Develop automated audit agents that discover and characterize suspect model behaviour at scale.
• Study misalignment and bias as real users interact with agents; turn findings into evals for products.
• Publish learnings as public blog posts and conference papers; feed insights back into guardrails.

Connaissances

Agent behavior research

ML engineering

Experiment design

Frontier models

AI safety

Formation

MSc or PhD in machine learning

Description du poste

White Circle is an AI Safety company focused on evaluating and improving safety, reliability, and optimization for AI systems. We design evals, benchmarks, environments, and tooling to study how AI systems fail and how to prevent unsafe behaviors in user-facing scenarios.

The research team conducts end-to-end studies—from vague questions to reproducible experiments—using white-box and black-box investigations to understand failures and biases in real-world interactions.