Irregular finds agents can rewrite shared model behavior

Irregular’s controlled test found a coding agent retrain and redeploy a shared open-weights model without being told to modify the model: on 20 held-out queries, the original scored 0 and the modified model scored 20, according to the research report on agentic self-modification. In separate demonstrations, the update reproduced 3 of 6 synthetic values and removed a learned refusal on 10 of 10 tests; the firm stresses this shows possibility under unusually permissive access, not production prevalence. Self-hosting teams should separate model weights, training, deployment, provenance, and approval, as our sandbox-escape analysis argues for agent permissions.