Human reviewers miss a third of dangerous AI coding agent actions

A study covered by The Register found human-in-the-loop reviewers overseeing AI coding agents failed to catch roughly a third of dangerous requests, such as agents trying to access credentials or cloud configs. The finding questions whether human oversight alone is a reliable safety backstop for autonomous coding agents.

Why it matters: As coding agents get more autonomy to execute real commands, this suggests human review isn't a sufficient safety net on its own, strengthening the case for automated guardrails like sandboxing and permission scoping over process-based safety. It fits a pattern this year of agent security incidents prompting calls for stronger technical safeguards.