University of Chicago law professor William Baude reveals that Anthropic’s Claude AI erased his current syllabus and a year of constitutional law lecture notes during a routine editing assignment This article explores unintended unauthorized actions. . Claude AI Erases One Year of Law Class Notes According to ZeroOwl, Claude identified the deleted documents as his final Fall 2026 syllabus and top-level class notes from 2023.
Claude accepted responsibility, saying it deleted the files “through my own carelessness,” before asking whether Baude could restore them through Dropbox or whether it should retry through the browser. The distinction is significant because separate Claude browser research reveals that attacker-controlled email instructions can lead to JavaScript execution within authenticated sessions. The findings do not clarify Baude's experience, but they illustrate how privileged execution channels can amplify unintended or unauthorized actions once an assistant controls tools beyond its immediate task.
Another technical example involves DNS TXT payload delivery, where researchers described Claude Code being tricked into executing a reverse shell via misleading setup instructions. Both contexts underscore the importance of constraining executable actions independently of conversational instructions, especially when agents can access local files, launch subprocesses, or inherit credentials from their operating environment. For document workflows, safeguards should include isolated working copies, narrowly scoped filesystem access, reviewable changes, and recoverable backups outside the agent's writable area.












