The safety of agents should never be done at the cost of speed or vice versa. Cerbere is the two
Most guardrails have only one layer: a model that guesses if a prompt is nasty. 19ms to 120ms, and many false positives. Cerberus is 3 layers: Layer 1 - Rules / Regex: 0.87ms It's the determinist guardrail. If the a
Most guardrails have only one layer: a model that guesses if a prompt is nasty. 19ms to 120ms, and many false positives.
Cerberus is 3 layers:
- Layer 1 - Rules / Regex: 0.87ms
It's the determinist guardrail. If the action is prohibited, it is blocked net. No guess, no latency. This is where we do 0.34ms in heat.
- Layer 2 - ML: 251ms
For cases that the rules do not cover. The model classifies the risk of the action.
- Layer 3 - LLM Judge: 362ms
Only for ambiguous cases. It's slow, but it's rare. We only pay the price of the LLM when we really need it.
Result: 95% of the actions are decided in <1ms, and really hard cases have a real judge.
We don't filter words. We control actions, with the right layer at the right time.
Code + bench: https://github.com/chrismsmr-celcom/cerbere-AG
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.