When you purchase through links on our site, we may earn an affiliate commission. Here’s how it works .
The developer frequently uses AI agents, and became a bit annoyed that a lot of them didn't clean up after themselves, leaving copious amounts of junk behind in the /tmp directory. Proving the old adage that once you have a hammer, everything looks like a nail, Guillemot asked Claude Fable to write a script that would sandbox each agent under its own folder in /tmp and run a cleanup after they were done. The main problem, naturally, was not deleting files that were actually in use.
Fable suggested adding logic to detect running agents and delay deleting their slice of /tmp, but Guillemot then told the bot the resulting code was far too complicated. Perhaps because the script involved hard deletion of data, Fable took it upon itself to perform an adversarial review, meaning the agent ran a new copy of itself to safety-check its own findings. Anthropic's harness then deemed the script risky enough to downgrade the model to Opus 5 and then Opus 4.8.
Guillemot stopped the process, but not in time. Adding insult to injury, after the agent wiped Guillemot's work, it did in fact leave /tmp behind. Some users suggested a few tools like Termaxa and other workarounds for these situations, but the fact that these tools need to exist in the first place is somewhat ironic.
Claude Opus 5 mistakenly deletes dev’s entire profile directory during routine backup, responds with 'Sorry, typo'
Anthropic's Claude hacked three real-life companies during security capabilities test
Anthropic warns Claude AI is building itself faster than expected, calls for option to halt frontier development
The fact that the harness dropped to a lower-end model due to safety concerns likely contributed to the problem, too. Given that Fable 5 outperforms Opus 4.8 in coding tasks, it might have caught the contradiction with the variable names in the test.
The developer got most of his data back thanks to collecting information from git, nix, session logs, and so forth. Yet there's some irony: all these agents running, and not one daily backup.
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
Key considerations
- Investor positioning can change fast
- Volatility remains possible near catalysts
- Macro rates and liquidity can dominate flows
Reference reading
- https://www.tomshardware.com/tech-industry/artificial-intelligence/SPONSORED_LINK_URL
- https://www.tomshardware.com/tech-industry/artificial-intelligence/claude-nukes-a-developers-700-gb-home-directory-while-testing-a-script-to-ensure-it-wouldnt-do-so-automatic-model-downgrade-may-have-contributed-to-the-screw-up#main
- https://www.tomshardware.com/membership
- Save a ridiculous 76% on this 7-in-1 Hyper Dual Monitor Thunderbolt 4 Dock — just $24 buys a dual 4K60Hz HyperDrive docking station for pennies on the dollar
- Samsung's new Odyssey G6 monitor hits a ridiculous 1,100Hz refresh rate at 720p — displays each frame for less than a millisecond
- Why Scaling AI Compute Performance Requires a New Power Architecture
- Save a ridiculous 76% on this 7-in-1 Hyper Dual Monitor Thunderbolt 4 Dock — just $24 buys a dual 4K60Hz HyperDrive docking station for pennies on the dollar
- NVIDIA Nemotron 3.5 Lightning and NeMo Switchyard Deliver Faster, Smarter, More Efficient Agentic AI
Informational only. No financial advice. Do your own research.