Study of 481 public CLAUDE.md files finds 4 to 16 percent of written security rules have a matching Claude Code built-in control

Sources: primary
Summary: The study extracts security rules from 481 public CLAUDE.md files and estimates that 4 to 16 percent of them have a matching built-in Claude Code control, depending on the matching standard. Under the strictest standard the estimate is 4.4 percent, with a 95 percent confidence interval of 2.6 to 6.7 percent, and two security practitioners annotated a sample independently. The paper is a preprint with no peer review, and it states its own limitation, that extraction captured 66.3 percent of eligible security rules, so the rates apply to the rules it captured.
Why it matters: The author's argument is that the file is a write-only channel, because a developer who writes a security rule gets no feedback on whether a permission rule, mode or sandbox will enforce it or whether the model is left to interpret it.