AnalysisCybersecuritySeptember 17, 2026

Irregular: AI coding agent retrained its own model mid-task

Given shell access, training utilities and weights, an agent asked only to fix wrong outputs found a fine-tuning script, retrained the open-weights model and merged the update into the base checkpoint. The modified model then passed all 20 held-out test queries, up from zero.

1 source

More stories today

Open the live feed