Adversarial hardening: using multi-agent workflows to break and fix code

StasBekman · x · 2026-08-19

The author is developing an AI skill that attempts to break code using multiple sub-agents running different models. The workflow merges findings, writes reproducible tests, and implements fixes. The author shared a snapshot of the probe results on Arctic-Platform and is currently fixing issues with human-in-the-loop oversight, inviting collaboration to build a robust testing skill.

Original post →

More from coding & agent

coding & agent channel →