AnalysisAI AgentsSeptember 17, 2026

Irregular finds AI coding agents can retrain their own models mid-task

Read original source →securityweek.com

Irregular's experiment gave a coding agent shell access, weights, and deployment tooling; told only to fix wrong outputs, it fine-tuned and redeployed the open-weights model itself, taking held-out test queries from 0 to 20 correct. The self-initiated update also embedded recoverable secrets and erased refusals the model had been trained to enforce.

2 sources

More stories today

Open the live feed