I’m building a free, open-source security layer for AI agents — and I want the community to help build it

Github

I’m building Cerbere AG, an open-source security and observability runtime for AI agents.
The idea is simple:
See what your agent does in production. Detect risky behavior. Block dangerous actions. Flag ambiguous ones. And require human approval when the system shouldn’t make the decision alone.
But this post isn’t really about announcing another AI security product.
It’s about making this a community project.
I want to make Cerbere AG free for everyone.
No credit card.
No paid plan required to start.
No artificial wall preventing developers from experimenting with agent security.
If you’re building an agent with tools, APIs, MCP, browser automation, databases, or other real-world capabilities, you should be able to add a security layer without first having to convince your company to spend thousands of dollars.
And I’m not pretending it’s finished.
It isn’t.
There are probably things I’ve missed.
There are edge cases I haven’t thought about.
There are attacks I haven’t covered.
There are integrations that could be better.
There are probably architectural decisions that someone in this community will look at and say:
“You should do this differently.”
Good. Tell me.
If you find a bug, report it.
If you find a security issue, report it.
If you think the detection logic is wrong, challenge it.
If you have a better approach, open a PR.
If you want to add an integration, build it.
If you have a new attack scenario, add it to the benchmark.
If you use Cerbere and discover that something doesn’t work, tell me.
The goal isn’t for me to personally build every feature.
The goal is to build something useful enough that other people want to contribute to it.
I’m building this largely by myself from Kinshasa, DRC, and I don’t have a large engineering team behind me.
So I can’t realistically cover every framework, every agent architecture, every tool integration, and every attack pattern alone.
That’s exactly why I’m opening it up.
What Cerbere currently focuses on
Agent execution tracing
Tool-call monitoring
Risk detection
Blocking
Flagging
Human approval for ambiguous/sensitive actions
Audit logs
Security evidence
Agent behavior observability
The project is still evolving, and I would rather have people tell me where it is wrong than pretend it is already perfect.
If you’re building AI agents, try it.
Use it.
Break it.
Audit it.
Contribute to it.
Tell me what sucks.
And if you think the idea is useful, give the repository a ⭐ or share it with someone building agents.
The more people test it, the more useful it can become.
Free. Open source. No credit card. Built in public.
GitHub: https://github.com/chrismsmr-celcom/cerbere-AG
If you see a problem, don’t just complain about it.
Bring the problem to the repo. Let’s build the solution together.

Comments

[Tera Tokomi]

Even if it is free, you still need to make sure someone elses work isnt being copied.