Safety Drift in Human-Centered Coding Agents: Suppression vs. Representation Shaping After Benign Adaptation

Modern coding agents retrieve repository context, plan edits, call tools, run tests, and negotiate changes with developers. Their safety is typically certified once, at release, through refusal behavior, policy prompts, decoding filters, tool permissions, and post-hoc vulnerability checks. We argue that this static view is insufficient. Deployed agents are routinely adapted through repository-specific tuning, LoRA adapters, organization-specific instruction or preference tuning, and tool-policy customization and such benign adaptation can make unsafe coding trajectories recoverable again even as task utility stays flat or improves. We call this safety drift. We distinguish suppression, which removes unsafe behavior at a narrow surface, from representation shaping, which reorganizes the agent's internal trajectory distribution so that safe behavior persists across a neighborhood of benign updates. We then propose adaptation-aware evaluation: measure safety, utility, tool use, and human-verifiability before and after benign customization, and report per-category retention. Durable, human-centered safety is a lifecycle property, not a release-time checkbox.