Challenge: Existing security benchmarks only cover user-level prompts and environmental threats . however, these models are vulnerable to pop-up attacks and prompt injections .
Approach: They propose a security benchmark that covers a set of six attack vectors that span both user-level and environment-level manipulations.
Outcome: The proposed security benchmarks cover a set of six real-world web environments with 2,970 adversarial trajectories and a multi-layered evaluation protocol dissecting agent failures across internal reasoning, behavioral execution, and task outcomes.

Similar Papers

VisualWebArena: Evaluating Multimodal Agents on Realistic Visual Web Tasks (2024.acl-long)

Copied to clipboard

Challenge: Existing benchmarks focus on text-based agents, neglecting many natural tasks that require visual information to effectively solve.
Approach: They propose a benchmark to assess the performance of multimodal web agents . they use visual and textual inputs to process and interpret natural language instructions .
Outcome: a new benchmark assesses the performance of multimodal agents on visually grounded tasks . the benchmark identifies limitations of text-only agents and offers insights towards building stronger agents for the web .
SecureVibeBench: Benchmarking Secure Vibe Coding of AI Agents via Reconstructing Vulnerability-Introducing Scenarios (2026.acl-long)

Copied to clipboard

Challenge: Existing benchmarks fail to capture scenarios in which vulnerabilities are introduced by humans . we evaluate 5 popular code agents supported by 5 LLMs on SecureVibeBench .
Approach: They propose a benchmarking tool that compares 105 C/C++ secure coding tasks . they use real-world open-source vulnerabilities and a comprehensive evaluation tool .
Outcome: The proposed benchmarks show that code agents struggle to produce correct and secure code . the best performing agent produces merely 23.8% correct and secured solutions .
MalURLBench: A Benchmark Evaluating Agents’ Vulnerabilities When Processing Web URLs (2026.findings-acl)

Copied to clipboard

Challenge: Existing models struggle to detect elaborately disguised malicious URLs, despite their ability to process malicious URL's.
Approach: They propose a benchmark to evaluate LLMs’ vulnerabilities to malicious URLs and a lightweight defense module to mitigate the vulnerability.
Outcome: The proposed framework analyzes 61,845 attack instances spanning 10 real-world scenarios and 7 categories of real malicious websites.
AutoPenBench: A Vulnerability Testing Benchmark for Generative Agents (2025.emnlp-industry)

Copied to clipboard

Challenge: LLM agents are promising for vulnerability testing, but lack benchmarks to evaluate and compare them.
Approach: They propose an open-source benchmark for the evaluation of vulnerability testing agents that includes 33 tasks ranging from introductory exercises to actual vulnerable systems.
Outcome: The proposed benchmark includes 33 tasks ranging from introductory exercises to actual vulnerable systems.
When "Correct" Is Not Safe: Can We Trust Functionally Correct Patches Generated by Code Agents? (2026.acl-long)

Copied to clipboard

Challenge: Code agents are increasingly trusted to autonomously fix bugs on platforms such as GitHub, yet their security evaluation focuses on functional correctness.
Approach: They propose to attack functionally correct yet vulnerable (FCV) patches by combining multi-turn reasoning with tool invocation and environment interaction.
Outcome: The proposed FCV-Attack achieves an attack success rate of 40.7% on GPT-5 Mini + OpenHands.
AgenticEval: Toward Agentic and Self-Evolving Safety Evaluation of Large Language Models (2026.findings-acl)

Copied to clipboard

Challenge: Existing static safety evaluation methods are ill-equipped to address dynamic nature of AI risks and evolving regulations, creating a critical safety gap.
Approach: They propose a new paradigm of agentic safety evaluation reframing evaluation as a continuous and self-evolving process rather than a one-time audit.
Outcome: The proposed framework shows a consistent decline in model safety as the evaluation hardens.
MCPEval: Automatic MCP-based Deep Evaluation for AI Agent Models (2025.emnlp-demos)

Copied to clipboard

Challenge: Existing evaluation frameworks suffer from limitations such as static task benchmarks, limited scope, and inadequate integration with practical applications.
Approach: They propose an open-source, Model Context Protocol-based evaluation framework specifically tailored for comprehensive and systematic assessment of LLM-powered agents.
Outcome: The proposed framework uncovers nuanced performance patterns and identify domain-specific strengths and weaknesses, providing valuable insights beyond traditional binary success metrics.
A Functionality-Grounded Benchmark for Evaluating Web Agents in E-commerce Domains (2026.acl-long)

Copied to clipboard

Challenge: Existing benchmarks focus on product search tasks, but ignore potential risks.
Approach: They propose a data generation pipeline that leverages webpage content and interactive elements to create diverse, functionality-grounded user queries.
Outcome: The proposed framework assesses the performance and safety of web agents under dynamic, real-world e-commerce environments.
MAPS: A Multilingual Benchmark for Agent Performance and Security (2026.findings-eacl)

Copied to clipboard

Challenge: Existing benchmarks do not provide a comprehensive, multi-domain, security-aware evaluation of multilingual agentic AI systems.
Approach: They propose a multilingual benchmark suite to evaluate agentic AI systems across languages and tasks.
Outcome: The proposed framework evaluates agentic AI systems across languages and tasks.
SALAD-Bench: A Hierarchical and Comprehensive Safety Benchmark for Large Language Models (2024.findings-acl)

Copied to clipboard

Challenge: SALAD-Bench is a safety benchmark specifically designed for LLMs . it provides a robust source for evaluating both attack and defense algorithms .
Approach: They propose a hierarchical safety benchmark specifically designed for LLMs . it uses a taxonomy of questions spanning three levels and a robust taxonomies based on a QA pair .
Outcome: The proposed safety benchmark shows that LLMs are resilient against emerging threats and the effectiveness of contemporary defense methods.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations