llm-safety

New method flags risky tool calls in AI agents before they happen

Researchers built a sparse-autoencoder and probe-based toolkit that reads AI agent internals to flag risky or unnecessary tool calls before execution happens.