Provoking LLMs by LLMs: Red Teaming for Harmful Content Detection
A red team framework for proactively detecting harmful content generated by large language models.
As large language models are deployed more broadly, understanding how they can be provoked into producing harmful content becomes essential to using them safely. This project investigates red teaming strategies for LLMs — using LLMs themselves as adversarial probes — to proactively identify vulnerabilities and harmful behaviors.
The work analyzes attack taxonomies, benchmarks, and security risks across prompt-based and model-level adversarial techniques, and contributes to evaluation frameworks aimed at improving the robustness, safety, and trustworthiness of generative AI systems (Jabbar et al., 2025).