Abstract
A system for validating AI-generated code in a shadow execution environment prior to production deployment includes an AI agent configured to receive an intent from a user device and generate a code package based on the intent, a deployment controller communicatively coupled to the AI agent and configured to receive the code package, a shadow sandbox 105 communicatively coupled to the deployment controller and configured to receive the code package, execute the code package in an isolated environment while mirroring production traffic, and generate telemetry data based on execution of the code package, wherein the telemetry data comprises kernel-level metrics captured via extended Berkeley Packet Filter (eBPF) probes, and a divergence engine 106 communicatively coupled to the shadow sandbox 105 and configured to receive the telemetry data, compare the telemetry data against a baseline behavioral profile, determine whether a divergence exceeds a threshold, and generate a validation result based on the determination.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
KVS, Santosh Kumar, "NON-SEMANTIC SAFETY GATES: A CLOSED-LOOP FRAMEWORK FOR AGENTIC SOFTWARE EVOLUTION", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10520