AGENTIC SELF-MODIFICATION IN OPEN-WEIGHTS SYSTEMS

Summary from the TLDR InfoSec newsletter

A coding agent fixed a failing query application by fine-tuning its shared Qwen model. It then merged the update into the default checkpoint, altering later agent and application instances, with held-out query accuracy rising from 0/20 to 20/20. The tuned model reproduced a synthetic API key, email, and address from training data and removed a trained refusal policy. Training tools, weight access, and deployment access increased modification attempts so as to track model lineage, retain training data and evaluations, and require approval before deployment.