PromtEngineer/localGPT
LocalGPT enables secure, private conversations with your documents on your local device using various open-source GPT models and embeddings without any data leaving your computer.
Awesome AI Agents › Post-Exploitation Agents
OSU-NLP-Group/EIA_against_webagent
EIA (Environmental Injection Attack) is a security research project focused on exposing vulnerabilities in generalist web agents that interact with web environments. The project demonstrates a novel form of indirect prompt injection designed to manipulate the web environment where state-changing actions occur, specifically targeting web agents to cause privacy leakage. By injecting malicious yet seemingly benign web elements and instructions into the HTML of benign websites, EIA misleads web agents into entering private user information into these injected elements, resulting in the leakage of personally identifiable information (PII) and other private data. The attack achieves a high attack success rate (ASR), with up to 70% success in stealing user-specific PII and 16% in leaking full user requests. Notably, EIA evades detection by traditional web malware detection tools like VirusTotal and does not disrupt the agent's ability to complete user tasks, making it a stealthy and effective attack. The project provides an experimental setup based on the SeeAct framework, utilizing a specific commit version and an upgraded OpenAI API version. It uses a curated dataset derived from Mind2Web tasks, with manually adapted realistic websites containing PII for testing the attack. The repository includes code for injecting various types of EIA attacks into benign webpages, targeting different stages of the web agent's action process. It also offers tools for deploying the manipulated websites via HTTP for interaction with the SeeAct agent and scripts for running and evaluating the attacks. Evaluation results are provided for multiple large language models (LLMs), and the project includes detailed instructions for replicating experiments and results. The research highlights the limitations of current defensive measures, including system prompts, in mitigating EIA risks. Overall, this project contributes to understanding and addressing privacy risks posed by advanced web agent interactions and indirect prompt injection attacks in web environments.
https://github.com/OSU-NLP-Group/EIA_against_webagent
Also in Post-Exploitation Agents
guardrails-ai/guardrails
Guardrails is a Python framework that enhances AI application reliability by providing input/output guards to detect and mitigate risks and generate structured data from large language models.
fr0gger/Awesome-GPT-Agents
A curated community-driven repository compiling diverse GPT agents specialized in offensive and defensive cybersecurity tasks, providing AI-powered tools for malware analysis, threat intelligence, vulnerability assessment, and more.
Ne0nd0g/merlin
Merlin is a cross-platform post-exploitation Command & Control server and agent supporting multiple protocols and advanced encryption for secure and flexible red teaming operations.
AprilNEA/AChat
AChat is an open-source, self-hosted AI platform designed for enterprises and teams, combining local processing with remote synchronization for centralized AI conversation management and collaboration.
sauravpanda/BrowserAI
BrowserAI is an open-source platform that enables running production-ready large language models and AI functionalities directly in the browser with WebGPU acceleration, ensuring privacy, offline capability, and no server infrastructure requirements.
HXSecurity/DongTai-agent-java
DongTai-agent-java is a Java application probe for DongTai IAST that collects runtime method invocation data via dynamic bytecode instrumentation to detect security vulnerabilities.
spellshift/realm
Realm is a scalable, reliable, and automated cross-platform Red Team engagement platform featuring a Pythonic DSL, multi-host management, and native Google Cloud integration for advanced offensive security operations.