Three reports from the past few days share one thing: AI is doing more and more work autonomously, and so the damage grows when things go wrong. A coding agent deleted an entire project in a minute and a half. The president of De Nederlandsche Bank expects AI fraud to hit here too. And cheap Chinese models are being used more and more. What does that mean for your work?
Claude Code deletes 48,218 files in 103 seconds
A developer reported on Reddit that a Claude Code agent deleted files from a live project on 21 September 48,218, including the Git history. Cybersecurity News wrote about it. The agent was meant to rebuild a copy of the project and wrote a Python script to clean up the old copy. That copy contained 614 Windows junctions pointing to the real project folder. The script did not recognise these junctions as links and, through them, also deleted the real files. The agent then admitted: "I broke something."
A caveat: the claim comes from a Reddit post with an attached report, not from independent forensic research. The post has since been removed, and the coverage does not mention any response from Anthropic.
The article's advice applies to any coding agent: treat it as an automated system with permissions, not as a conversation partner. Run deletion actions as a dry run first, move files rather than deleting them outright, and let the agent work in a sandboxed folder with as few permissions as possible. If you work in software development, make sure your backups sit beyond the agent's reach. For how Claude handles your data, see the model page.
DNB: "It is simply going to happen"
Olaf Sleijpen, President of De Nederlandsche Bank, said on Sunday on WNL Op Zondag that AI, alongside cyber risks, ranks among the top three risks for the financial sector. The trigger was the fraud at the Italian bank Intesa Sanpaolo, where scammers posed as the chief executive on WhatsApp and used AI to imitate the voice of a lawyer. The haul ran into the tens of millions of euros. Sleijpen does not rule out a comparable case in the Netherlands.
Do you work in financial advisory or as an accountant? Then agree that any urgent request to transfer money is always verified through a second, pre-agreed channel, even if you recognise the voice or the sender.
Chinese models: cheaper and increasingly used
According to a Mozilla report, Chinese open-weight models lag behind the leading closed American models by an average of about 4.4 months, and cost up to five times less per task, Android Headlines reports. You can see this in the OpenRouter ranking that we retrieve daily. On 26 September, DeepSeek accounted for 21.32% of all tokens, GLM for 8.34%. ChatGPT reached 11.63% and Claude 2.4%.
Low costs say nothing about your data. DeepSeek's service stores data in China and by default uses it for training. GLM operates internationally through a Singaporean entity. Both models are open-weight, however, so you can run them yourself or with a European provider. For sensitive data, that is the only defensible route. For what applies to each model, see /veiligheid/.
What you can do now
- Only give coding agents write permissions in a sandboxed working folder, and keep backups out of their reach.
- Set up a call-back procedure for payment requests, separate from the channel the request came in through.
- Choosing a cheap model? First compare on /modellen/ where your data ends up.