🚨 Claude Code deleted 48,218 files in 103 seconds
According to a developer’s post on Reddit, the AI agent was supposed to rebuild a mirrored copy of the project. While clearing the old mirror, it ran a Python script that, through 614 Windows directory junctions (special links that pointed to the real working project directories), began deleting files already in the original tree.
📁 As a result, 48,218 files were deleted. Git was also affected: parts of ".git/objects", "refs", and "logs" were lost, making it harder to restore the history.
After that, the agent reported:
“Craig — stop and read this. I broke something”.
⚠️ The main lesson: a dangerous combination of AI autonomy, file system access, and the ability to execute commands without additional confirmation.
For AI agents, critical measures include sandboxing, minimum privileges, dry-runs, confirmation for dangerous operations, and independent backups.
Important: the incident is described by the developer themselves and was not independently verified by Anthropic.
#claudecode #Anthropic #Aİ #AIAgents #CyberSecurity
According to a developer’s post on Reddit, the AI agent was supposed to rebuild a mirrored copy of the project. While clearing the old mirror, it ran a Python script that, through 614 Windows directory junctions (special links that pointed to the real working project directories), began deleting files already in the original tree.
📁 As a result, 48,218 files were deleted. Git was also affected: parts of ".git/objects", "refs", and "logs" were lost, making it harder to restore the history.
After that, the agent reported:
“Craig — stop and read this. I broke something”.
⚠️ The main lesson: a dangerous combination of AI autonomy, file system access, and the ability to execute commands without additional confirmation.
For AI agents, critical measures include sandboxing, minimum privileges, dry-runs, confirmation for dangerous operations, and independent backups.
Important: the incident is described by the developer themselves and was not independently verified by Anthropic.
#claudecode #Anthropic #Aİ #AIAgents #CyberSecurity
