Reproduce “LatentTriad: Pre-Action Risk Representations for Malicious Agent Skill Detection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhe Yu, Yi Liu, Ying Zhang, Simin Chen
LatentTriad screens task-and-skill packages before an agent generates output, invokes tools, or executes skill code. It reads the model’s pre-action states across the skill’s purpose, operations, and task context, then applies a risk detector. On Large-ID, it reaches 99.0% pooled and 94.4% source-macro malicious-class F1 on Llama-3.1-8B; source holdout lowers source-macro F1 to 73.2%, so screening does not replace runtime enforcement.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.