Purpose
Agentic Security is an open-source vulnerability scanner for agentic workflows and large language models (LLMs), and it can also serve as an AI red teaming toolkit. Its goal is to protect AI systems from threats such as jailbreaks, fuzzing, and multimodal attacks, helping developers, researchers, and security teams proactively identify and mitigate risks in AI systems, thereby ensuring safer and more reliable deployments.
Key Capabilities
- Multimodal attacks: Probe vulnerabilities across text, image, and audio inputs to ensure that LLMs can withstand diverse threats.
- Multi-step jailbreaks: Simulate complex iterative attack sequences to discover weaknesses in LLM security mechanisms.
- Comprehensive fuzzing: Stress-test any LLM with random inputs to identify edge cases and unexpected behavior.
- API integration and stress testing: Seamlessly connect to LLM APIs and push them to their limits through high-volume, real-world attack scenarios.
- Reinforcement learning-based attacks: Use reinforcement learning to build adaptive, intelligent probes that evolve as model defenses evolve.
Usage
Install the package via pip to get started with Agentic Security.
Agentic Security uses plain-text HTTP specifications, for example:
<<PROMPT>>
where <<PROMPT>> will be replaced with the actual attack vector during scanning. Insert the Bearer XXXXX header value to fill in your application credentials.
Add an LLM integration template
Adding LLM integration templates is supported.
Add your own dataset
To add your own dataset, you can place one or more CSV files containing the prompt column, and these data will be loaded when agentic_security starts.
Run as a CI check
A default configuration example is provided that can be run as a CI check.
Extend the dataset collection
Add new metadata to agentic_security.probe_data.REGISTRY and implement a loader in it.
Dynamic datasets and mutation
One of the current examples uses sampling from existing preloaded prompt data and applying mutations to generate new datasets.
In the custom integration example, use /v1/self-probe for integration testing. This endpoint randomly simulates a fake LLM refusal response.
To probe the image modality, you can use the following HTTP request:
Replace XXXXX with your actual API key, and<<BASE64_IMAGE>> is the image variable.
To probe the audio modality, you can use the following HTTP request:
Replace $GROQ_API_KEY with your actual API key, and ensure that the file parameter points to the correct audio file path.
CI/CD integration
The provided example GitHub Action is designed to perform automated security scans. The example GitHub Action workflow ensures that project security is maintained in a continuous integration manner.
Module The class is used to manage prompt processing and interaction with external AI models and tools. It supports asynchronously fetching, processing, and publishing prompts to detect model vulnerabilities. See module.md.
For more detailed information on how to use Agentic Security (including advanced features and customization options), please refer to the official documentation.
Applicable Scenarios
- Developers, researchers, and security teams conduct security assessments of LLMs and agentic workflows.
- Integrate automated security scanning into CI/CD pipelines.
- Through custom datasets and integration templates, perform customized red team testing for specific application scenarios.
Limitations and Notes
- Dates in the roadmap and future goals are tentative and may be adjusted based on project progress and priorities.
- The project welcomes contributions; please read the contribution guidelines before contributing.
- Agentic Security is released under the Apache License v2.
- The project declares that it has no association whatsoever with cryptocurrency, blockchain technology, or related initiatives: no tokens, no tokens, only code.
