Agentic Security
An open-source vulnerability scanner for LLMs and agent workflows that probes for jailbreaks, fuzzes inputs across text, image and audio, and can run as a CI check to catch regressions in an AI system's safety guardrails.
Agentic Security is a vulnerability scanner for large language models and the agent workflows built on them, aimed at developers, researchers and security teams who need to test their own AI systems for weaknesses rather than exploit someone else's. It probes multimodal inputs across text, images and audio, runs multi-step iterative jailbreak sequences to uncover gaps in an LLM's safety mechanisms, stress-tests an API with high-volume randomized fuzzing, and can use reinforcement learning to craft adaptive probes that evolve against a target model's defenses.
A target is configured with a plain HTTP request template, such as a chat-completions call, with a placeholder that the scanner substitutes with each attack vector during a run, and custom prompt datasets can be added as CSV files with a prompt column, loaded automatically at startup. It ships a web UI alongside its CLI, and an init command sets up configuration for running scans as an automated check in continuous integration.
It installs via pip and is released under the Apache-2.0 license.