Safety Testing LLM Agents at Scale: From Risk Discovery to Evidence-Grounded Verification

Yunhao Feng; Ruixiao Lin; Ming Wen; Qinqin He; Yanming Guo; Yifan Ding; Yutao Wu; Jialuo Chen; Zhuoer Xu; Xiaohu Du; Jianan Ma; Zixing Chen; Xingjun Ma; Yunhao Chen; Xinhao Deng

arXiv · 2607.01793

Safety Testing LLM Agents at Scale: From Risk Discovery to Evidence-Grounded Verification

Yunhao Feng·Ruixiao Lin·Ming Wen·Qinqin He·Yanming Guo·Yifan Ding·Yutao Wu·Jialuo Chen·Zhuoer Xu·Xiaohu Du·Jianan Ma·Zixing Chen·Xingjun Ma·Yunhao Chen·Xinhao Deng

View original source ↗Read source document ↗Export metadata ↓

Abstract

LLM agents increasingly perform autonomous actions through external tools, leading to complex and evolving safety risks. However, existing safety testing targets expert-designed safety violations, and the corresponding outcomes are evaluated by hard-coded rules, making them costly to extend as agents evolve. To this end, we present Vera, an end-to-end automated safety testing framework that instantiates software engineering testing principles for non-deterministic agents through a three-stage, self-reinforcing pipeline. First, a literature-driven exploration continuously discovers and structures emerging risks into taxonomies of safety risks, attack methods, and tool execution environments. Second, combinatorial composition across taxonomy dimensions produces executable safety cases, each specifying a concrete safety goal, a programmatically constructed initial state, and a deterministic verification predicate grounded in observable artifacts. Third, adaptive execution runs heterogeneous agents in isolated sandboxes where a control agent steers multi-turn interaction based on runtime observations, while evidence-grounded verifiers judge outcomes from environment state and tool-call evidence rather than model self-report. We evaluate Vera on four production agent frameworks (OpenClaw, Hermes, Codex, Claude Code), revealing substantial safety weaknesses, with average attack success rates reaching 93.9\% under multi-channel attacks; we also release Vera-Bench, comprising 1600 executable safety cases spanning 124 risk categories across three execution settings. These results indicate that modular, executable testing infrastructure is essential for rigorous and maintainable safety evaluation of rapidly evolving agentic systems at scale. The code is publicly available at https://github.com/Yunhao-Feng/Vera.

Explore related subjects

cs.AI

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Yunhao Feng, Ruixiao Lin, Ming Wen, Qinqin He, Yanming Guo, Yifan Ding, Yutao Wu, Jialuo Chen, Zhuoer Xu, Xiaohu Du, Jianan Ma, Zixing Chen, Xingjun Ma, Yunhao Chen, Xinhao Deng. 2026-07-04. Safety Testing LLM Agents at Scale: From Risk Discovery to Evidence-Grounded Verification. https://arxiv.org/abs/2607.01793

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Collab-Solver: Collaborative Solving Policy Learning for Mixed-Integer Linear Programming

Mixed-integer linear programming (MILP) has been a fundamental problem in combinatorial optimization. Conventional MILP solving mainly relies on carefully designed heuristics embedded in the branch-and-bound framework. Driven by the strong capabilities of neural networks, recent research is exploring the value of machine learning alongside conventional MILP solving. Although learning-based MILP methods have shown great promise, existing works typically learn policies for individual modules in MILP solvers in isolation, without considering their interdependence, which limits both solving efficiency and solution quality. To address this limitation, we propose Collab-Solver, a novel multi-agent-based policy learning framework for MILP that enables collaborative policy optimization for multiple modules. Specifically, we formulate the collaboration between cut selection and branching in MILP solving as a Stackelberg game. Under this formulation, we develop a two-phase learning paradigm to stabilize collaborative policy learning: the first phase performs data-communicated policy pretraining, and the second phase further orchestrates the policy learning for various modules. Extensive experiments on both synthetic and large-scale real-world MILP datasets demonstrate that the jointly learned policies significantly improve solving performance. Moreover, the policies learned by Collab-Solver have also demonstrated excellent generalization abilities across different instance sets.

cs.AI↗

arXiv2026

Fact Grounded Attention: Eliminating Hallucination in Large Language Models Through Attention Level Knowledge Integration

"The greatest enemy of knowledge is not ignorance, it is the illusion of knowledge." Large Language Models have conquered natural language but remain prisoners of their own probabilistic nature--confidently hallucinating facts they never truly knew. We present Fact Grounded Attention (FGA), a novel architectural modification that transforms unreliable language models into deterministic truth tellers by injecting verifiable knowledge directly into the attention mechanism. Unlike existing approaches that patch hallucinations after generation or prepend retrieved text, FGA intervenes at the mathematical heart of the transformer--the pre-softmax attention scores--creating a model that cannot hallucinate when facts exist in its knowledge base. Our experiments across 1,107 technical queries spanning smartphones, laptops, and electric vehicles demonstrate a transformation from 6.3% accuracy in vanilla Llama 3.2 to 99.7% accuracy with FGA. More critically, knowledge updates occur in under one second without retraining, compared to hours for parameter editing approaches. FGA doesn't just reduce hallucination--it eliminates it entirely for verifiable facts, marking a fundamental shift from probabilistic approximation to deterministic precision in neural language generation.

cs.AI↗

arXiv2026

MemeLens: Multilingual Multitask VLMs for Memes

Memes are a dominant medium for online communication and manipulation because meaning emerges from interactions between embedded text, imagery, and cultural context. Existing meme research is distributed across tasks (e.g., \textit{hate, misogyny, propaganda, sentiment, humour}) and languages, which limits cross-domain generalization. To address this gap, we propose \textsc{MemeLens}, a unified multilingual, multitask explanation-enhanced Vision-Language Model (VLM) for meme understanding. We consolidate $38$ public meme datasets, filter and map dataset-specific labels into a shared taxonomy of $20$ tasks spanning harm, targets, figurative/pragmatic intent, and affect. We present a comprehensive empirical analysis across modeling paradigms, task categories, and datasets. Our findings suggest that robust meme understanding requires multimodal training, varies substantially across semantic categories, and remains sensitive to over-specialization when models are fine-tuned on individual datasets rather than trained in a unified setting. We make the experimental resources (https://github.com/MohamedBayan/MemeLens), model (https://huggingface.co/QCRI/MemeLens-VLM) and datasets (https://huggingface.co/datasets/QCRI/MemeLens) publicly available to the community.

cs.AI↗