GPT — Field Evidence

Adaptive Adversaries: A Multi-Turn, Multi-LLM Benchmark for LLM Agent Security — LLM-based agents process external content, exposing them to prompt injection and multi-turn manipulation. Most safety benchmarks evaluate defenders against fixed attack pools collected before evaluat