This is a research prototype. The data and analyses are preliminary and not yet validated — we'd welcome your .
An AI coding assistant tasked with software repairs deleted over 48,000 live files and destroyed a Git repository database by following Windows junction pointers instead of treating them as directory links.
A Claude Code user reportedly authorized an AI coding agent to repair isolated copies of software used to analyze historical stock-options data. A sub-agent allegedly followed Windows directory junctions into the live project tree, deleting about 48,218 files and the Git object store in 103 seconds. The account has not been independently verified.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
AI systems that fail to perform reliably or effectively under varying conditions, exposing them to errors and failures that can have significant consequences, especially in critical applications or areas that require moral reasoning.
Also classified as
The user allowed an autonomous AI agent direct write/delete access to a live local environment without maintaining a remote backup, representing unsafe deployment and overreliance.
The AI system failed to perform reliably under standard operating conditions because it lacked the capability to recognize and handle Windows junction folders correctly, leading to catastrophic data loss.
AI system
Due to a decision or action made by an AI system
Unintentional
Due to an unexpected outcome from pursuing a goal
Post-deployment
Occurring after the AI model has been trained and deployed
The file deletion was caused by the AI system's failure to correctly interpret Windows junction pointers, which was an unexpected and unintentional outcome of its programming task.
An AI coding assistant unintentionally deleted 48,218 files and corrupted a local Git repository on a developer's machine due to a failure to correctly interpret Windows junction pointers. This is a localized software reliability and data loss incident with negligible implications for national security.
The AI system was deployed as a coding agent tasked with executing a series of complex, multi-step repair jobs autonomously. It made decisions on how to clean directories and execute commands without step-by-step human intervention, maintaining an internal workflow until it failed.
Minimal or No Risk: The AI system is a general-purpose coding assistant used for personal software development and data analysis, which does not fall under prohibited or high-risk categories and carries no additional regulatory obligations.
The AI system was deployed as a coding assistant tasked with performing software repairs and rebuilding a testing environment.
The report provides a clear, technically detailed account of the failure mechanism (Windows junctions) and the exact scale of the data loss (48,218 files). The secondary anecdote about OpenClaw is also straightforward.
The incident could be interpreted purely as human error due to the developer's lack of remote backups, but the direct cause of the deletion was the AI's failure to handle directory pointers.
Technology
AI inability to recognize junctions
The AI treated Windows directory junctions as normal folders and followed them.
Rapid automated file deletion
The AI deleted over 48,000 live files in under two minutes without validation.
Data Inputs
Misconfigured testing environment
The testing environment contained 614 junctions pointing to live files.
Human Factors
Lack of remote repository backup
The developer did not push the project to a remote server like GitHub.
Inadequate system isolation
The AI was allowed to run directly on a system with live working files.
Process and Methods
Failure to use version control
No robust local or remote version control was active to restore lost history.
Unrestricted write access
The AI coding agent was given direct write/delete access to the filesystem.
Management
Overreliance on AI capabilities
The user trusted the AI to manage file copies safely without oversight.
AI system involved
Identified by automated extraction; can differ from the alleged developers above.
