Claude's security framework flaw leads to 700gb data deletion

Developer Sebastien Guillemot reported on August 29 that Claude mistakenly deleted approximately 700GB of his data, including his entire user home directory, during a test of an AI agent's file deletion protection mechanism. The incident occurred after Anthropic's security system automatically downgraded the model to Opus 4.8 due to perceived risk. Guillemot had requested a script to clean up temporary files generated by AI agents, but the security environment flagged the task as high-risk, leading to multiple model downgrades. During testing, the system incorrectly matched deletion targets with the user's home directory, resulting in the accidental deletion. Despite efforts to recover data through Git and logs, the incident highlights the risks of AI agents handling high-risk file operations, even when safety checks are in place. The error stemmed from a logical conflict between test and cleanup code using the same variable name, which was not detected by the model. Guillemot noted that a more capable version of the model might have identified the issue, but this remains speculative. The incident underscores the need for robust safeguards when AI systems manage critical file operations, as even well-designed security measures can fail due to internal logic errors.