Hey Product Hunt 👋 Guardana actually started as something we needed for our own work. While building and running our own AI models and AI-powered systems across different projects, I kept looking for a security framework that we could adapt to the way we actually build. I didn’t want just another fixed scanner or a large collection of jailbreak prompts. I wanted something we could use as a security framework around the whole development lifecycle: — a solid set of security checks out of the box — the ability to add our own rules for project-specific risks — custom evaluators when generic grading simply isn’t enough — support for different models, agents and infrastructure — something we could run locally during development — put into pytest and CI/CD as an actual security gate — and keep using against deployed systems afterwards Most importantly, we needed to be able to adapt the security logic to the actual application. Because a healthcare agent, an internal company assistant, a coding agent and an AI system with access to production tools do not have the same threat model. That became Guardana. Today Guardana is an Apache-2.0 open-source AI security verification framework covering multiple layers of an AI system: 🛡️ Build-time security Scan model files, repositories, dependencies, prompts, templates and supply-chain risks. 🎯 Runtime verification Probe live models and agents for prompt injection, jailbreaks, unsafe behavior and tool-related vulnerabilities. 🔌 MCP security Verify MCP servers and parts of their authorization and tool surface. 🔎 Execution trace analysis Analyze what an AI system actually did — model calls, tools, retrieval, identities and scopes, approvals, memory and side effects. ↔️ Security regression testing Compare releases and detect when a change to a model, prompt, tool or deployment makes the system less secure. 🧩 Extensibility Create your own rules, evaluators, targets and private security packs without having to fork the project. And the same engine can be used from the CLI, pytest, CI/CD or scheduled health checks. One design principle has remained important from the beginning: “Couldn’t verify” must never silently mean “secure.” If an evaluator cannot determine the outcome, a capability is missing, the target is unavailable or the test cannot execute correctly, Guardana keeps that uncertainty visible instead of quietly producing a green check. We built Guardana because we wanted this flexibility for our own AI projects. Now we’re opening it up because I suspect many other teams are running into the same problem: generic security checks are useful, but sooner or later every serious AI system needs security rules that understand its own architecture, permissions, tools and business logic. Guardana is: — Apache-2.0 — no account required — no mandatory telemetry — no mandatory cloud — designed for private and self-hosted environments — extensible with your own security logic We’ve already shipped 14 small releases and we’re still moving quickly. I’d especially love feedback from people building real AI systems: What project-specific security rule or evaluator would you need before you could use something like Guardana as a real deployment gate? And if you’re running self-hosted models, agents, MCP servers or your own AI infrastructure, I’d love to hear where Guardana still falls short.

Guardana

Customizable security gates for AI systems

Customizable security gates for AI systems