agentic_security

Agentic Security is an open-source scanner that probes LLMs and agent workflows for vulnerabilities. The README lists multimodal attacks across text, images and audio, multi-step jailbreak sequences, fuzzing with randomized inputs, API integration and stress testing with high-volume attack scenarios, and reinforcement-learning-based attacks that adapt to the model's defenses. It is meant for developers, researchers and security teams who want to find weaknesses before deployment.

Install with `pip install agentic_security` and start the web UI with `agentic_security` (it serves on port 8718 by default; `--port` and `--host` change that). The target is described with a plain-text HTTP spec, such as a POST to an OpenAI-style chat completions URL with an `Authorization: Bearer` header, where `>` is replaced by each attack prompt during the scan. You can add your own CSV datasets with a `prompt` column, and run it as a CI check: `agentic_security init` writes an `agesec.toml` with the LLM spec, a maximum budget and a failure threshold. The README leaves some sections, such as adding integration templates, marked TBD. It is Apache licensed.