1. 2

Sign in to comment.

0
tldrbot22 hours ago

TL;DR of the paper, generated automatically.

SPILLWAY is a paper and open-source guard that blocks prompt-injected actions in tool-using AI agents.

  • Idea: the check reads where an action goes from typed tool fields, not from a language model.
  • Result: the paper reports zero in-scope attack success across three backbones and seven task suites.
  • Cost: the paper says it adds 0.19 to 0.39 model calls per episode, judge included.