Reproduce “When Tools Silently Lie: Evaluating and Mitigating Blind Compliance in Tool-Augmented Data Agents” 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
Zifu Tao, Changqing Yin
ToxicBench tests whether data-analysis agents recover when successful tool returns are silently altered while source data stays fixed. It pairs clean and poisoned conditions and tracks both evidence checks and adopted answers. In 118 tasks, GPT agents’ task success fell by 26–39 percentage points under poisoning; the benchmark uses constructed tables and does not establish production prevalence or transfer.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.