AI SAFETY · RED TEAMING · MODEL EVALUATION
Caroline V. James
AI red teamer and model evaluator focused on adversarial testing, safety, and model behavior.
SELECTED WORK
ABOUT
The hardest model failures rarely happen in a vacuum. They emerge from context: ambiguous intent, social and cultural cues, competing instructions, hidden assumptions, and information a model understands imperfectly or too literally. I bring a social-science perspective to AI safety and red teaming, with an emphasis on designing tests that expose those failures.
EXPERTISE
Adversarial testing across safety domains, from direct attacks and hard boundary cases to indirect prompt injection and hidden-instruction attacks. Evaluation work includes groundedness, bias, reasoning, safety behavior, synthetic data, persona and dialogue development, and the translation of qualitative model behavior into structured evaluation criteria. Cultural and subcultural context, coded language, social dynamics, and ambiguous intent are recurring areas of specialization.
Adversarial prompt design & red teaming · Indirect prompt injection / XPIA · Model evaluation & safety analysis · Synthetic data & dialogue · Persona & use-case development · Qualitative research