Claude Downgrade Bug Deletes Developer's 700GB Home Directory
Developer Sebastien Guillemot lost 700GB of data after Anthropic's safety framework automatically downgraded Claude from Fable 5 to Opus 4.8 due to perceived risk. The downgraded model generated a buggy cleanup script that bypassed safety checks and deleted the user's home directory. This incident highlights a critical flaw in AI safety guardrails, where automatic model downgrades intended to mitigate risk can lead to lower-quality code generation that causes catastrophic data loss. It underscores the dangers of allowing AI agents to execute high-risk file operations without strict human oversight. The data loss occurred because the generated script reused the same variable name for both the test logic and the cleanup logic, causing it to delete the home directory while ironically leaving the target `/tmp` directory intact. Although the developer recovered most data using Git and Nix, the event exposes how logic errors in test code can bypass path-filtering safety checks.
## BACKGROUND
Large language models (LLMs) often use safety guardrails to filter inputs and outputs to prevent harmful actions. In some AI agent environments, when a task is flagged as high-risk, the system may automatically downgrade the active model to an older or more restricted version. However, older models like Claude Opus 4.8 may lack the advanced reasoning capabilities of newer models like Fable 5, increasing the likelihood of generating buggy code.