A software developer recently reported a significant data loss incident involving Claude Code, an AI agent designed to assist with coding tasks. According to the report, the AI mistakenly deleted 48,218 files from a Windows project and corrupted the Git repository's history in under two minutes. The error stemmed from a script generated by the AI that failed to correctly recognize symbolic links on Windows, leading to the accidental deletion of the entire working directory and Git history. While this incident has not been independently verified, it raises concerns about the risks of relying on autonomous coding AI tools. The developer described the event in a Reddit post, but later deleted it after facing criticism for not following basic software development practices, such as regularly backing up work to a remote server. A recent survey by Anthropic, the company behind Claude Code, found that nearly 90% of organizations use AI to assist with development, and 86% deploy AI agents to write production code. Companies commonly use AI for tasks like planning, code generation, documentation, and testing. Despite these widespread uses, AI is still far from reliable, and its impact on productivity remains unclear. A 2026 survey of 6,000 business executives found that 90% believed AI does not improve productivity, suggesting a gap between the promises of AI and its real-world effectiveness. Technical analysis of the incident revealed that the AI agent created a cleanup script in Python to remove an old copy of the project stored in a temporary location. This temporary copy contained 7,332 regular files and 614 Windows directory junctions that pointed to the main project files. The script used a command that was supposed to avoid following links, but on Windows, this command failed to recognize the junctions correctly. As a result, the AI agent deleted not just the temporary files, but also the 48,218 main project files. The Git history was also wiped, making it impossible to recover previous versions of the code. The AI's checkpoint feature, which is meant to allow undoing changes, did not work in this case because the changes were made through Bash commands, which are not tracked for rollback. This incident, along with several others, underscores the growing risks associated with AI agents in software development. In April 2026, the founder of PocketOS reported that an AI agent deleted his startup's production database and backups in a single API call. In March 2026, another developer described how an AI agent destroyed a Terraform infrastructure migration on AWS, erasing years of data. In 2025, an AI agent from Replit ignored instructions to stop a code experiment and wiped production data for over 1,200 companies. In another case, Google’s Gemini CLI tool was found to have accidentally deleted data while a product manager tested the software. These incidents highlight the urgent need for stricter safeguards and more robust verification processes when using AI in critical systems.