Agentic Security: Open-Source LLM Vulnerability Scanner
Agentic Security is an open-source tool designed to identify vulnerabilities in Large Language Models (LLMs) and agent workflows. It focuses on detecting jailbreaks, prompt injections, and multimodal attack vectors across text, images, and audio inputs. The tool is specifically tailored for developers securing chat and agent endpoints, providing a structured approach to red-teaming and security testing.
Developed as a community-driven project by Alex Miasoiedov, Agentic Security operates under the permissive Apache 2.0 license and is implemented in Python. Its primary purpose is to automate red teaming efforts, offering a comprehensive suite of attack simulations to uncover weaknesses in AI systems before deployment.
How It Works
Agentic Security functions by fuzzing API endpoints with predefined attack datasets. The tool supports custom datasets in CSV format, enabling users to extend its attack libraries with their own prompts or mutations.
During operation, Agentic Security loads attack modules—such as jailbreak datasets, multimodal probes, or reinforcement learning-based attacks—and systematically tests an LLM’s resilience. Results are compiled into failure rate reports per attack module, with configurable thresholds (low, medium, high) to classify vulnerabilities.
The tool logs detailed scan results, including duration, module-specific failure rates, and status indicators. For continuous integration, Agentic Security provides CLI commands (`agenticsecurity ci`) to automate security scans within CI/CD pipelines, such as GitHub Actions.
Key Features & Capabilities
- Multimodal Attack Testing: Evaluates vulnerabilities across text, image, and audio inputs. Audio probes use multipart/form-data requests with files (e.g., `.m4a`) and models like `whisper-large-v3`, while image inputs follow similar HTTP-based workflows.
- Multi-Step Jailbreak Simulations: Emulates iterative attack sequences to expose flaws in LLM safety mechanisms.
- Reinforcement Learning (RL)-Based Attacks: Experimental support for adaptive attack vectors, where an attacker LLM dynamically evolves strategies based on model responses. This feature is part of the roadmap for future expansion.
- API Integration & Stress Testing: Directly interfaces with LLM APIs (e.g., OpenAI, Groq) using standard HTTP methods. Users replace placeholder tokens (e.g., `Bearer XXXXX`) with actual API keys and specify endpoints like `/v1/self-probe` for testing.
- Custom Dataset Support: Users can add new datasets by extending the `ProbeDataset` class or registering metadata in the `REGISTRY`.
- CI/CD Integration: Designed for automated security checks, Agentic Security includes CLI commands for initialization (`agenticsecurity init`), module listing (`agenticsecurity ls`), and continuous scanning (`agenticsecurity ci`). Sample workflows for GitHub Actions demonstrate how to embed scans into development pipelines.
Use Cases
Agentic Security is primarily used by developers and security teams to preemptively identify vulnerabilities in AI-driven applications.
- Chatbot & Agent Security: Testing endpoints for jailbreaks, prompt injections, and adversarial inputs before deployment.
- Red Teaming Automation: Simulating real-world attack scenarios to benchmark LLM defenses. The tool’s modular design allows teams to focus on specific threats (e.g., multimodal exploits) without manual effort.
- Continuous Security Validation: Integrating into CI/CD pipelines to enforce security gates.
- Research & Benchmarking: Evaluating new LLM models or safety mechanisms against standardized attack datasets. Researchers can compare failure rates across models or attack types to inform defensive strategies.
The tool’s open-source nature and permissive licensing encourage collaboration, with plans to expand dataset coverage (e.g., 100,000+ prompts) and introduce community-contributed modules for broader threat coverage.