Unintended Agents · Issue 1 · 10–16 November 2026
What AI agents broke this week, and the setting that would have stopped it.
Every week we read the public bug trackers for Claude Code, Codex and OpenClaw, plus the Cursor forum, for one kind of report: an agent did something its user never wanted, and something real broke. No attackers, no jailbreaks. Just agents doing their jobs badly.
This week we found 16 reports of material harm. Over the past 30 days we found 87. Almost none of them appear in any incident tracker or news story.
Here are the five that teach the most.
1. An overnight subagent wiped a Windows data drive
What happened. A user left a multi-agent session in the Claude desktop app running overnight without approvals. A subagent wanted a copy of a project inside WSL and ran rsync -a --delete with a shell variable as the destination. The command passed through wsl.exe, an extra shell layer expanded the variable to nothing, and the destination became /. rsync deleted everything the user could delete, including business documents, design files and months of backups on the mounted D: drive. Files deleted through WSL skip the Recycle Bin.
Two details make it worse. Stopping the subagent ended wsl.exe on the Windows side, but rsync kept running inside WSL for about two more minutes. And the subagent's first damage report said the documents folder was "not touched, checked". It had only looked for error lines, and rsync does not log successful deletes.
The fix.
- In any script an agent writes, a variable that feeds a delete must fail when empty:
"${DEST:?}/", not$DEST/. - If you run agents unattended on Windows, turn off WSL's drive automount (
[automount] enabled = falsein/etc/wsl.conf). The agent then cannot reach your Windows drives from Linux at all. - After a destructive command goes wrong, do not trust the agent's damage report. Check the files yourself.
Source: claude-code #109735 · Evidence: B (detailed self-report) · Report likely drafted by the agent · Mode: unattended overnight
2. An "empty" database schema took a production app down for two hours
What happened. A user approved an agent's plan to drop a legacy schema from their production Supabase project. The agent had checked that its tables held zero rows. It had not checked that Supabase's API layer (PostgREST) still listed the schema as exposed. Once the schema was gone, every API request from the user's live mobile and web app returned 503. The user saw an empty app and believed all their projects were deleted. No data was lost, and recreating the empty schema fixed it.
This is the one case this week that another tracker (Eon) also logged.
The fix.
- "Empty" is not "unused". Before dropping anything, the agent should search config for references to it, not just count rows.
- In Supabase, remove a schema from Settings → API → Exposed schemas first. Drop it a day later.
- Run destructive migrations on a branch or staging project before production. Your approval is only as good as the plan you were shown.
Source: claude-code #109368 · Evidence: B · Report likely drafted by the agent · Mode: user approved the deletion
3. An agent spent the app's whole API balance testing a prompt
What happened. A developer asked Claude Code to improve one prompt in their production app. The agent edited the prompt, pushed it live, and then wrote a script that sent 100 sample inputs to the live endpoint. It ran the full test seven times in a row without asking. The balance hit −$0.04, and the app's AI features went offline for its users. The agent's own words afterwards: "I ran the 100 test notes through Sonnet 7 times in a row without asking you and without checking the balance."
The money was small (about $2.90). The outage for real users was not.
The fix.
- Never give a coding agent your production API key. Give it a separate key in its own workspace, with a low spend cap.
- Tell the agent to test against a mock or a staging endpoint, and put that endpoint in the project config so it does not have to guess.
Source: claude-code #108897 · Evidence: B · Written by the user · Mode: auto-accept edits
4. An agent broke a rate limit it had agreed to, and the provider banned the account
What happened. A user told an agent to stay within a paid compiler API's limits: 30 requests a day, 5 an hour. The agent wrote these limits into its own plan. Later it sent 50 requests in one hour, 45 of them back to back, on the theory that refused requests "are not counted". The provider disabled the user's paid account. The agent read the "Compilation is disabled for this account" error as an ordinary compiler error and sent 25 more requests. The provider refused a refund, citing a policy violation.
The fix.
- A limit written in a prompt or a plan is a suggestion. A limit written in code is a limit. Wrap the API in a small script that counts requests and refuses past the cap, and let the agent call only the wrapper.
- Make any account-level error (disabled, suspended, quota) stop the whole run.
Source: claude-code #110248 · Evidence: B · Report likely drafted by the agent · Mode: not stated
5. A voice agent sent an email with the one thing it was told to leave out
What happened. A user dictated an email through Codex desktop's voice mode and told it explicitly not to mention an earlier third-party update. The sent email contained a paragraph summarizing that update. Before the send, the voice assistant had said several times that the email was "sent and verified" while it was still a draft with missing attachments. After the send, it promised to remove the excluded text, though the email had already gone.
The fix.
- Let an agent draft emails. Send them yourself, after reading the final text.
- An agent's check that the email "was sent" says nothing about whether it says the right thing.
Source: codex #61378 · Evidence: B · Report likely drafted by the agent · Mode: user authorized the email
The pattern this week: the guard has to live outside the agent
In two of these cases the agent knew the rule and broke it anyway. It had written the rate limit into its own plan. It had been told what to leave out of the email. In the other three, nobody thought to state a rule at all, because the risk looked obvious to a human.
Either way, the fix that works is the same, and none of it depends on the agent remembering anything: a spend cap on the key, a request counter in a wrapper, a drive the agent cannot reach, a send button only you press.
Also this week
- An agent filled a system drive with about 1 TB of temporary builds and maxed the CPU. (claude-code #109176, evidence C)
- An agent's own script deleted about 540 layout records, then the agent misreported the count and the cause. (claude-code #109022)
- In auto mode, an agent worked on the wrong branch despite repeated instructions: it filed 42 issues, merged PRs and stashed the user's files. (claude-code #109269)
- A Codex agent with approvals off kept expanding its diagnostics past every timebox and burned 1.18 billion tokens with nothing delivered. (codex #61834)
- Overnight parallel agents used about 67% of a user's weekly Max plan allowance. (claude-code #110418)
- A Cursor agent wrote a GitHub token in plain text into git remotes and shell history in six places. (Cursor forum)
How we pick and grade cases
We count a case only when an agent decided to take the harmful action. Pure product bugs do not count. We read every report in full and grade the evidence:
- A: the vendor engaged with the report.
- B: a detailed self-report with commands, logs or steps someone could check.
- C: a thin report, or one that rests on the agent's own account.
No vendor has confirmed any case in this issue. Many reports are written or drafted by the agent itself, and we say so on each one. We also note the permission mode, because most serious cases happen when people turn oversight down.
Seen an agent do something you never asked for? Reply to this email.