This is a research prototype. The data and analyses are preliminary and not yet validated — we'd welcome your .
A user granted an Anthropic Claude agent permission to organize a desktop and delete temporary files. The agent mistakenly targeted and deleted a folder containing 15 years of family photos using terminal commands, which bypassed the system trash. The files were eventually recovered via an iCloud feature after standard recovery methods failed.
Venture capitalist Nick Davidov alleged that Anthropic's Claude Cowork, after being asked to organize his wife's desktop and delete temporary Office files, instead deleted a folder containing roughly 15 years of family photos and related personal memories via terminal commands. Davidov later reported that Apple support helped recover the files through an iCloud recovery feature.
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 granted the AI agent direct write and delete permissions on a live file system, representing an unsafe use of a system not yet robust enough for such tasks.
The AI system failed to perform reliably, mistakenly targeting and deleting a critical folder of photos instead of temporary files, which represents a lack of capability and robustness in executing file system tasks.
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 incident was caused by an action made by the AI system, which mistakenly deleted a folder of photos while pursuing the user-defined goal of organizing the desktop.
Added categories peak at 2: Minor (Psychological Harm, Direct)
An Anthropic Claude AI agent mistakenly deleted 15 years of family photos on a user's personal desktop. While distressing for the individual, the incident is a localized consumer safety and software reliability issue with negligible national security implications.
The report describes the AI system as an 'agent-based AI workspace' designed to 'delegate multi-step tasks like organizing files, data analysis, and document drafting' and 'reading, creating, and editing files directly on a user's computer'. It executed a sequence of actions (renaming, deleting via terminal) autonomously to achieve the goal of organizing the desktop.
Risk Level 3: Limited Risk. The system is an agent-based AI workspace operating within a chatbot interface ('Claude desktop interface'), which falls under the category of chatbots requiring transparency obligations.
The AI system is Anthropic Claude, a general-purpose chatbot and assistant, which was deployed here as an agent to perform file organization and content curation tasks on a user's desktop.
The report is a first-hand account from the user detailing exactly what happened, the specific AI system involved (Anthropic Claude agent), the actions it took (deleting files via terminal), and the outcome (recovery via iCloud).
None.
Technology
Terminal Execution Bypass
AI executed deletion via terminal, bypassing the standard system Trash.
Flawed File Renaming Logic
AI attempted to rename files but mistakenly deleted the entire folder.
Lack of Safety Safeguards
Agent lacked safeguards to prevent modifying critical personal directories.
Data Inputs
Ambiguous Directory Context
AI misidentified the photo folder as a temporary office file directory.
Lack of File Classification
No clear metadata distinguishing temporary files from critical user photos.
Human Factors
Excessive Permission Granting
User granted write and delete permissions to the AI on a live file system.
Lack of Immediate Backup
User initiated file cleanup without ensuring a robust local backup existed.
Process and Methods
Direct File System Modification
Process allowed direct modification of files without a safe dry-run preview.
Non-Reversible Deletion Method
System used terminal commands that bypass the standard OS trash recovery.
Management
Premature Feature Release
Releasing file-system agent features before they are fully safe for mainstream.
Regulatory Environment
Lack of AI Agent Standards
Absence of safety regulations for autonomous AI agents manipulating files.
AI system involved
Identified by automated extraction; can differ from the alleged developers above.
