Aug 2026· 2026 IEEE 34th International Requirements Engineering Conference Workshops (REW)· pp. 155-158· 0 citations· 12 references
Abstract
Crowd-Based Requirements Engineering (CrowdRE) increasingly relies on multi-step large language model (LLM) pipelines to elicit, classify, and prioritize requirements from large volumes of user-generated content. Unlike prior batch natural-language-processing (NLP) tools, these agentic pipelines maintain persistent memory across sessions, invoke external tools, and pass intermediate representations through chains of model calls-each transition a potential exfiltration surface for contributor data. To our knowledge, no existing CrowdRE framework, RE methodology, or data-protection guidance document specifically addresses the class of privacy failures that emerges at the intersection of agentic AI architecture and crowd-based feedback pipelines. This problem statement characterizes four novel failure modes-cross-user memory bleed, implicit preference fingerprinting, tool-chain exfiltration, and synthetic crowd laundering-and presents an architectural threat model that exposes where each failure manifests in a representative agentic CrowdRE pipeline. We pose five open research questions and call on the community to treat contributor privacy as a first-class quality attribute before agentic CrowdRE becomes the production norm.
This work presents LIDAR (LLM Identification from Decisions and Actions at Runtime), an active black-box fingerprinting method for coding-agent execution that achieves high Top-1 accuracy and MRR and outperforms four existing fingerprinting and API-auditing baselines.
Chu-Yi Wang, Xiao-Hui Xie, Tong-Ze Wang et al.· 2 citations· ⚡1
Pincer is a new defense that operates at the resource layer and works alongside existing defenses at the tool-call layer like the auto mode, an isolated-context model that automatically learns and enforces dynamic user-specific least-privilege policies.
Mayank Rathee, Alexander Stepanov, Shalin Madabhavi et al.· 0 citations
Computer-use agents (CUAs) are agents powered by vision-language models (VLMs) that perceive a screen and operate an operating system through mouse, keyboard, and terminal interactions to automate everyday digital tasks. Their exposure to untrusted content creates a risk of indirect prompt injection (IPI), where an adv...
Chen Xiong, Zhi-Yuan He, Pin-Yu Chen et al.· 0 citations
AgentHop, a diagnostic benchmark of 1,011 multiple-choice questions paired with a controlled seven-tool sandbox under fixed token, turn, and tool-call constraints, and the full benchmark set and the harness are released to support diagnostic agent benchmarking.
Chanhee Park, Jeongho Yoon, Sung-Bin Han et al.· arXiv.org· 0 citations
DarKnight, a self-improving agentic framework designed to run on a single workstation GPU inside an investigator's Tor crawler and solve CAPTCHAs with small, locally hosted open-weight VLMs that no provider decision can withdraw, without per-market training or third-party services is presented.
Bing-Hui Wu, Yeo Zi Yi, Darren Tan et al.· 0 citations
AGENTQ is proposed, an attack framework that combines layer-banded LoRA injection with partial-PGD repair over a multi-codebook quantization-equivalence class that preserves normal agentic capability while concentrating malicious behavior in the quantized model, underscoring the need to make quantization-aware safety e...
Xiao-Qun Liu, Qi-Ben Yan· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.