AI Safety Research Scientist: Trace & Defuse Agent Failures

Whitecircle; Paris; Hybride; EUR 90 000 - 140 000; Plein temps

Avantages offerts par ce poste

Paid time off

Hybrid Paris w relocation

Medical insurance

Equity package

Hardware & tools

Subscriptions for AI agents

Résumé du poste

White Circle is seeking a research scientist to study how LLM agents fail in the wild, eliciting deception and unsafe behaviours, and building a deeper understanding of misalignment in realistic user scenarios.

You will own end-to-end research from unclear questions to falsifiable experiments, develop audit agents at scale, and publish findings to inform guardrails and product safety. Hybrid Paris-based role with relocation support potential.

Qualifications

• Track record of empirical research in agent behaviour, model evaluation, alignment, or closely adjacent area.
• Able to build a research MVP involving fine-tuning, agent inference, and evals without platform team support.
• Experience designing experiments under real conditions: isolating failures, calibrating baselines, distinguishing signal from artifacts.

Responsabilités

• Own research projects end to end from unclear concerns to falsifiable experiments and defendable results.
• Develop automated audit agents that discover and characterize model behavior at scale.
• Study misalignment and bias as users interact with agents and turn findings into evals for products.
• Pressure-test frontier agents in realistic, high-stakes scenarios to find failure points before customers.
• Run white-box and black-box investigations to understand how AI models fail.
• Publish learnings as public blog posts and conference papers and feed back into guardrails.

Connaissances

Empirical research

ML engineering

Experimental design

AI safety

Frontier models

Formation

MSc or PhD in ML/CS/ cognitive science

Outils

Fine-tuning

Agent inference

Eval frameworks

Description du poste

White Circle is seeking a research scientist to study how LLM agents fail in the wild, eliciting deception and unsafe behaviours, and building a deeper understanding of misalignment in realistic user scenarios.

You will own end-to-end research from unclear questions to falsifiable experiments, develop audit agents at scale, and publish findings to inform guardrails and product safety. Hybrid Paris-based role with relocation support potential.