A new arXiv preprint examines why coding agents often take defensive actions that go beyond what a task requires. The authors point to how these models are post-trained: with outcome rewards, and with recent methods that concentrate credit on the specific turns that decide the outcome. The problem, they argue, is that over-defense escapes both mechanisms.
The root cause appears at what the paper calls "silent forks"—points where the agent makes a choice without any explicit signal that a decision is being made. Because these forks are silent, neither the outcome reward nor the credit-assignment method can properly penalize unnecessary defensive behavior. The result is that such actions persist in the final code.
The study is based on a single source, so no independent comparison is available. Still, the finding suggests a blind spot in current training approaches: if defensive actions are invisible to the reward signal, they will remain uncorrected even as other aspects of agent behavior improve.