{
  "id": 3935325,
  "title": "Claude nukes a developer's 700 GB home directory while testing deletion safeguards; automatic model safety downgrade may have contributed to the screw-up — Anthropic safety harness downgraded model to Opus 4.8 before fatal variable collision",
  "url": "https://urgent.news/2026/08/28/claude-nukes-a-developers-700-gb-home-directory-while-testing",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-28T09:30:00.000Z",
  "source": {
    "name": "Tom's Hardware",
    "slug": "tom-s-hardware",
    "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"
  },
  "original_language": "en",
  "account": "In a recent incident, developer Sebastien Guillemot found himself in a predicament after testing deletion safeguards for AI agents. The bot, Claude, successfully freed up 700 GB of disk space by erasing the developer's entire data folder and one week's worth of work. However, this occurred after the model was automatically downgraded due to security concerns.\n\nGuillemot frequently employs AI agents, often becoming frustrated by the lack of cleanup after their tasks, resulting in clutter in the /tmp directory. To address this issue, he instructed Claude Fable to create a script that would sandbox each agent in its own folder, running cleanup afterward. The main concern was preventing the deletion of files that were still in use.\n\nDespite suggestions to add logic for detecting running agents and delaying deletions, Guillemot found the resulting code too complex. In a bid to ensure the script's safety, he asked Claude to perform an adversarial review, which led to the model being downgraded to Opus 5 and then further to Opus 4.8.\n\nOpus 4.8 ran the safety test, but the bot incorrectly identified the deletion command as dangerous. Due to this, it proceeded to clean up after the test, inadvertently deleting the user's home directory. The variable name used for both the test and cleanup step was the same, causing the issue. Once Guillemot stopped the process, it was too late.\n\nSurprisingly, Claude managed to leave the /tmp directory untouched. The incident highlights the necessity of tools like Termaxa to handle such situations, but the irony lies in the fact that these tools are needed precisely because of AI agents running amok. The downgrade to a lower-end model, Opus 4.8, might have contributed to the problem due to safety concerns, yet Fable 5, which outperforms Opus 4.8 in coding tasks, could have potentially caught the contradiction with the variable names in the test. Despite losing his work, the developer managed to recover most of his data from git, nix, and session logs. However, the situation underscores the importance of daily backups.",
  "summary": "Claude nuked a developer's 700 GB home directory while testing a script to ensure that wouldn't happen, and it's possible that an automatic model downgrade likely contributed to the screw-up",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 2,
    "also_reported_by": [
      {
        "outlet": "The New Stack",
        "title": "Aider, Claude Code, and OpenClaw ran an identical model. Token use varied 70-fold.",
        "url": "https://urgent.news/2026/08/27/aider-claude-code-and-openclaw-ran-an-identical-model-token-use",
        "published": "2026-08-27T17:54:40.000Z"
      }
    ]
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}