Anthropic's Claude AI wipes developer's 700GB home directory during safety test
A developer using Claude AI agents asked the model to write a cleanup script for temporary files, but during an adversarial review, the AI deleted his entire home directory, losing 700 GB of data and a week of work. The incident occurred after Anthropic's safety harness automatically downgraded the model from Opus 5 to Opus 4.8, which then misidentified the deletion targets. The developer had previously noted that many AI agents left junk in /tmp, prompting the cleanup script request.
The deletion occurred during a self-initiated adversarial review, where the agent spawned a copy of itself to validate its own cleanup script. Anthropic's safety harness then automatically stepped in, downgrading the model from Opus 5 to Opus 4.8 after judging the script risky. The lower-tier model reused a single variable name for both the test's target-matching logic and the subsequent cleanup phase, causing the fatal collision.
Guillemot recovered most of his lost work by piecing together data from git repositories, nix store entries, and session logs. Notably, the agent did successfully delete the /tmp junk it was originally asked to clean, but left the directory itself intact afterward. The developer had no daily backup system running despite relying heavily on multiple AI agents.
This incident could heighten caution among developers who increasingly delegate system-level tasks to AI agents, potentially slowing adoption of autonomous coding tools in production environments. If safety mechanisms like automatic model downgrades can introduce new failure modes, organizations may demand more rigorous testing and human oversight before trusting agents with destructive operations. Individual developers could face significant data loss and productivity setbacks, while the broader industry may need to reconsider backup practices and sandboxing strategies as AI autonomy expands.