Awesome AI Agents › Runtime Analysis Tools
The project "agent-attack" is the official code and data repository for the paper titled "Dissecting Adversarial Robustness of Multimodal LM Agents," presented at ICLR 2025 and orally at NeurIPS 2024 Open-World Agents Workshop. This research focuses on evaluating and understanding the adversarial robustness of multimodal large language model (LM) agents, which integrate visual and textual information to perform tasks. The repository provides tools to reproduce the experiments and attacks described in the paper, including adversarial attacks on images that challenge the robustness of these multimodal agents. The project requires the use of a companion repository called VisualWebArena, which has been modified to support perturbations to trigger images, essential for the episode-wise evaluation of the agents. VisualWebArena is a large-scale environment requiring significant disk space and Docker for deployment. The repository includes scripts to run different types of adversarial attacks, such as captioner attacks and CLIP attacks, targeting various models like GPT-4V, Gemini, Claude, and GPT-4o. These attacks generate adversarial examples saved in an experiment data directory. The project also supports detailed evaluation methods, including episode-wise and stepwise evaluations, to assess the impact of adversarial perturbations on agent performance. Users must configure API keys for OpenAI, Anthropic, and Google Cloud services to run the code. The repository provides comprehensive setup instructions, including environment setup, API key configuration, and data processing. The project is significant for researchers and developers interested in the security and robustness of multimodal AI agents, offering a valuable resource for studying adversarial vulnerabilities and improving model defenses.
https://github.com/ChenWu98/agent-attack
PrivateGPT is a fully private, offline-capable AI platform that enables users to interact with their documents using Large Language Models through a comprehensive API and user-friendly interface, ensuring data never leaves the user's environment.
PentAGI is a fully autonomous AI-driven penetration testing system that integrates professional security tools, multi-agent AI collaboration, and scalable microservices architecture to deliver comprehensive and automated security assessments.
Impacket is a Python library offering low-level programmatic access and implementations for various network protocols, focusing on security research and educational use.
VIPER is a versatile and powerful red team platform that supports adversary simulation and cybersecurity assessments across multiple operating systems, enhanced with AI-driven automation and extensive post-exploitation capabilities.
MCP-Scan is a security tool that statically and dynamically scans MCP connections to detect and prevent vulnerabilities like prompt injections, tool poisoning, and cross-origin escalations, offering real-time monitoring and customizable guardrails for enhanced protection.
OpenRecall is a fully open-source, privacy-focused digital memory tool that captures and makes searchable your digital history through screenshots, supporting Windows, macOS, and Linux with local data storage and advanced semantic search capabilities.
MCP server that exposes Cheat Engine memory reading, scanning, disassembly and debugging to AI agent clients through roughly 180 tools over a named pipe or TCP relay.
Amon is a modern, open-source server monitoring platform designed to help users track and manage server performance and uptime effectively.