I’m building Cerbère-AG, a security evidence layer for AI agents.

Christopher dikesa — Sat, 26 Sep 2026 19:27:56 +0000

I’m building Cerbère-AG, a security evidence layer for AI agents.

Most AI security tools focus on what goes into the model: prompt injection, malicious inputs, jailbreaks, etc.

I’m focusing on what happens after the model decides to act.

Cerbère-AG observes and controls agent actions across tool calls, including:

tool-call monitoring and traces
policy enforcement
sensitive-action detection
argument and capability checks
budgets and execution limits
trajectory-level risk detection
human approval for sensitive actions
security evidence for AI agent activity

The idea is simple:

Don’t just ask whether an agent is safe to talk to. Ask whether it is safe to let it act.

I’m looking for developers and teams running AI agents in real or realistic environments to test Cerbère-AG and tell me where it fails.

I’m especially interested in design partners who can give real-world feedback on agent workflows, policies, approvals, and failure cases.

If you build AI agents, security tooling, MCP integrations, or autonomous workflows:

→ Give me your feedback: what would you expect a production-grade agent security layer to catch that Cerbère currently doesn't?

And if you find the project useful, a ⭐ on GitHub helps people discover it.

I’m more interested in breaking it and finding its weaknesses than in compliments.

If you have an agent that you think could expose a real failure mode, send it my way.

→ Test it:
→ GitHub:

We benchmarked our prompt-injection detector against OWASP's LLM Top 10

Christopher dikesa — Sat, 26 Sep 2026 19:11:17 +0000

We just published our first public benchmark for AgentGuard, the runtime we're building to secure AI agents in production.
The honest version: our deterministic (regex) layer alone catches 91.5% of prompt injection attempts with zero false positives, in under 2ms. Adding a ML layer pushes recall to 98.1% — but the trade-offs are real: ~450ms latency, and a 33% false-positive rate on benign prompts that were deliberately worded to look like attacks.
We're publishing the numbers as they are, weaknesses included, and mapped everything to the OWASP Top 10 for LLM Applications so it's easy to compare.

DEV Community: Christopher dikesa

I’m building **Cerbère-AG**, a security evidence layer for AI agents.

We benchmarked our prompt-injection detector against OWASP's LLM Top 10

I’m building Cerbère-AG, a security evidence layer for AI agents.