---
type: "post"
id: "claude-and-the-con-law-notes-incident-1df498"
slug: "claude-and-the-con-law-notes-incident-1df498"
title: "Claude and the Con Law Notes Incident"
description: "William Baude says a Claude Code agent accidentally deleted a year of constitutional-law lecture notes and his current syllabus while he was preparing to teach."
retrieval_nugget: "William Baude says a Claude Code agent accidentally deleted a year of constitutional-law lecture notes and his current syllabus while he was preparing to teach. Dropbox and other backups made the files easy to recover, but the account shows how a routine reading task can expand into risky script execution and why recoverability matters when agents work on local files."
published_at: "2026-11-20"
updated_at: "2026-11-20"
record_date: "2026-11-20"
date_kind: "published_at"
topics: ["claude","claude-code"]
entities: ["Claude Code","William Baude","Divided Argument","Dropbox"]
related_patterns: ["agent-security-moves-to-runtime-boundaries"]
novelty_label: "new_fact"
what_changed: null
prior_refs: []
source_url: "https://blog.dividedargument.com/p/claude-and-the-con-law-notes-incident"
source_urls: ["https://blog.dividedargument.com/p/claude-and-the-con-law-notes-incident"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:claude-and-the-con-law-notes-incident-1df498"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/claude-and-the-con-law-notes-incident-1df498/","markdown":"https://newruntime.com/posts/claude-and-the-con-law-notes-incident-1df498.md","json":"https://newruntime.com/posts/claude-and-the-con-law-notes-incident-1df498.json"}
markdown_url: "https://newruntime.com/posts/claude-and-the-con-law-notes-incident-1df498.md"
---

# Claude and the Con Law Notes Incident

## Retrieval answer

William Baude says a Claude Code agent accidentally deleted a year of constitutional-law lecture notes and his current syllabus while he was preparing to teach. Dropbox and other backups made the files easy to recover, but the account shows how a routine reading task can expand into risky script execution and why recoverability matters when agents work on local files.

## Summary

William Baude says a Claude Code agent accidentally deleted a year of constitutional-law lecture notes and his current syllabus while he was preparing to teach. Dropbox and other backups made the files easy to recover, but the account shows how a routine reading task can expand into risky script execution and why recoverability matters when agents work on local files.

## What happened

In a November 19, 2026 account on Divided Argument, law professor William Baude describes asking his teaching assistant, Claude, to check that page numbers in his notes were up to date. He returned to find that the agent had deleted a year’s worth of lecture notes and his current syllabus. Baude says the notes were easy to recover from a Dropbox backup and that he had another set of backups, so no lasting harm was done.

Baude specifies that he was using a Claude Code agent, not Claude’s website chat interface. He says he had relied on Claude Code for tasks including updating page numbers, downloading and organizing research, finding old files and fixing a printer. In his view, the agent’s ability to act on files makes it more useful for such work, but also gives it greater capacity to cause damage.

His account of the cause is qualified: “as I now understand it,” Claude appears to have decided it needed to write a bespoke script each time it read Word files. It then ran an optional cross-check across other lecture notes, which Baude says was intended to reduce omissions and mistakes. He believes the script was misnamed and that Claude did not understand how PowerShell worked. The result was deletion, not an intentional response to his instructions. Baude says he has since fixed the behavior that led Claude to write a new script to read Word files, but does not detail the change.

Baude frames the incident as a learning experience rather than an argument against using AI. He considers three possible responses: avoid agentic AI or confine it to a small sandbox; learn enough about agents to anticipate their failure modes; or continue using them while accepting trial and error and maintaining backups. He says he expects to learn most from the last approach. The account does not identify a Claude Code version or provide a reproducible procedure for triggering the deletion.

## What the archive adds

New Runtime’s earlier coverage includes [Claude Code’s experimental Code Mode and its controls around tool calls, memory, MCP calls and hooks](https://newruntime.com/posts/claude-code-changelog-anthropics-claude-code-776d38/) (November 20, 2026), and [a Claude Code mod for setting session spending limits](https://newruntime.com/posts/claude-code-mods-mods-for-claude-code-714c94/) (November 17, 2026). Those notes concern product capabilities, not this failure. Baude’s account is a new incident and adds a concrete example of accidental file deletion during agent work.

## Trend signal

This incident supports the broader concern behind [agent security moving to runtime boundaries](https://newruntime.com/patterns/agent-security-moves-to-runtime-boundaries/): limiting what an agent can change and retaining recoverable copies can matter more than relying on instructions alone. The account does not establish which boundary or sandbox would have prevented this particular failure, so it is an illustration, not evidence that a specific control is sufficient.

## Why it matters

File operations are a common part of useful agent work, yet a task that sounds like checking page numbers can lead to scripts, broader cross-checks and unintended writes. Teams should treat an agent’s ability to read and modify files as an operational permission, not as a harmless extension of chat. The lesson is not simply to keep agents away from important work: Baude emphasizes the capabilities he would lose by sharply limiting them.

For builders, the practical distinction is between containing damage and preventing every error. A constrained workspace can reduce the range of files an agent can affect; versioned or independently stored backups can make mistakes reversible. Neither control guarantees safety, and this account does not compare them. But the recovery described here depended on backups that existed before the failure, not on the agent recognizing its own mistake.

The account also cautions against evaluating file-capable agents solely by whether they can complete a requested task. The execution path matters: an agent may create auxiliary scripts or inspect a wider set of files while attempting verification. Teams need to decide in advance which directories and file types are safe for autonomous writes, and which must remain outside an agent’s working boundary.

## Limits and uncertainty

The account is a single user’s retrospective description. Baude presents the explanation as his current understanding; the specific script, naming error, PowerShell behavior and deletion sequence are not documented in enough detail for independent reproduction. He does not name a Claude Code version, describe the change he made to prevent bespoke Word-reading scripts, or report whether the behavior was investigated by Anthropic. No broader incident rate or comparison of sandboxing and backup strategies is established. The recovery outcome is specific to this user’s backups and does not show that other users’ files would be recoverable.

## Skill to build now

Map one agent-enabled workspace this week before granting it broad file access. Separate a disposable working directory from valuable originals, identify which folders the agent may write to, and confirm that an independent backup can restore a changed file. For tasks that may prompt scripts or broad file searches, start with copies rather than canonical documents. This does not eliminate unintended actions, but it turns recovery from an improvised response into a tested part of the workflow.

## Reader takeaway

An agent can turn a routine file task into consequential execution, even when nobody intended it to delete anything. If you let an agent work on valuable files, make recovery a designed part of the workflow: keep separate backups and consider limiting its writable workspace. Baude’s experience shows that backups can contain harm; it does not show that backups alone make agent access safe.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- topic hub: [Claude Code](https://newruntime.com/topics/claude-code.md): Explore the claude-code topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
