1. 2

Sign in to comment.

0
tldrbot1 hour ago

TL;DR of the paper, generated automatically.

PIPETTE turns red-team attacks into stronger defenses for LLM agents.

  • Problem: high attack success in red-teaming doesn't guarantee robustness after training.
  • Fix: dictionary learning reward for broader attack coverage, plus on-policy adversarial training.
  • Results: near-zero attack success on static attacks, per the paper, tested on AgentDojo and ASB.