An agent in 'ask first' mode moved 11GB without asking, a user says
Follow-up to Tuesday Nov 10: Your coding agents and your keys: Claude Code secrets hit the news, ChatGPT Desktop questionedThis week we covered Claude Code leaking secrets through its 'Don't Ask Again' option. Now a user reports the opposite problem: asking was on, and it still did not ask. They filed it on GitHub last Friday. This is one person's report, not confirmed.
They used Claude Code on Windows in manual approval mode, with a rule that PowerShell commands must stop and ask. They say they never saw a prompt. The agent ran 9 PowerShell commands that moved 11GB of files, and the log shows approvals they say they never gave.
The same day, a forum post claiming "Humble Bundle just killed most of their security through vibecoding" got 382 points on Hacker News. Vibecoding means letting AI write code with little checking. Our sources give no details. United Airlines now cites cybersecurity risks to block AI agents from booking flights. And Anthropic's new Claude Meetings joins Google Meet calls to take notes, on by default for Team accounts.
An approval button only protects you if it really works, and if the log tells the truth. Check your agents today: what can they touch, and do your logs match what you saw? Demand fast answers when they don't.
