The authors propose a two-dimensional perturbation space: four interaction targets (user utterances, procedural specifications, carried-over context, tool contracts) crossed with three corruption types (missing, redundant, incorrect), yielding 12 families.
LLM agents degrade reliably under controlled interaction perturbations
A systematic framework varying four input targets and three corruption types shows robustness gaps across ten models on two benchmarks.
Chinese Tech
Yuanjian Xu · Jinyi Han · Jianing Hao · Yuxuan Sun · Fanyang Lu · Richeng Xuan · +2 more
The Hong Kong University of Science and Technology (Guangzhou) · Tencent Hunyuan · East China Normal University · University College London
Research Digest··2 min read
Thread:Agent Security & Attacks
The authors introduce AgentPerturb, a framework that applies missing, redundant, and incorrect edits to user utterances, procedural specifications, carried-over context, and tool contracts in agentic tasks.
Why this paper
From Tencent Hunyuan and 3 others · Part of Agent Security & Attacks, now 94 papers
In one line
Perturbations to LLM agent inputs consistently degrade performance, and agents rarely detect corrupted information.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§