Cerbere-AG – The layer that ensures security, observability and control, which sits between the AI agent and the real world

Description and use

Cerbere-AG provides security, observability, and control for autonomous AI agents with three layers—Regex, ML, and an LLM Judge—reducing attack success and offering an audit trail.

Problem

AI agents act autonomously and can perform unintended or malicious actions, requiring guardrails that balance autonomy with control.

Outcome

Cerbere-AG reduces attack success to below 30% with Regex, achieves 100% recall with ML, and resolves ambiguous cases with an LLM Judge, while providing an audit trail for compliance.

Workflow

• 1. Agent issues a tool call.
• 2. Regex layer filters obvious malicious patterns.
• 3. ML layer classifies remaining calls, raising recall to 100% in tests but flagging uncertain cases.
• 4. LLM Judge evaluates ambiguous cases, considers context, and decides to allow, block, or flag for human approval.
• 5. Collector records spans and traces into an audit trail independent of the agent.

Evidence

Regex alone lets more than 30% of attacks through.

Source

ML pushes recall up to 100% in some of our tests, but raises false positives.

Source

The LLM Judge resolves the ambiguous cases ML can’t classify.

Source