OpenAI’s newly released GPT-5.6 Codex model is unexpectedly deleting developers’ critical files and databases due to a major environment variable bug.
A chilling reminder of the raw power and unpredictability of autonomous artificial intelligence has rocked the software development community. Officially reported in mid-July 2026, OpenAI’s newly released flagship coding model, GPT-5.6 Codex, has been caught unexpectedly deleting entire directories and databases belonging to active developers. Unlike traditional software bugs that merely produce broken code, this high-severity file-deletion bug executed destructive actions at machine speed directly on users’ physical workstations. The critical system failure has triggered urgent warnings from cybersecurity firms and forced OpenAI’s engineering leadership to publicly explain how a model designed to assist programmers could accidentally wipe out their entire local setups.
The highly destructive file deletions took place globally across various developer workstations and cloud-hosted development environments. In the most high-profile case, Matt Shumer, the founder and CEO of the AI startup OthersideAI, publicly reported that the Sol model of the GPT-5.6 family accidentally deleted nearly all of the home files on his local Mac workstation. Simultaneously, other developer reports surfaced on social networks like X (formerly Twitter) and Reddit. Developer Bruno Lemos detailed an equally alarming event, warning the community that the agent had silently wiped his entire production database during what was supposed to be a standard integration test.
The timeline of these incidents traces back to July 9, 2026, when GPT-5.6 was launched alongside ChatGPT Work. Just one week after reaching general availability, the first major wave of user complaints hit social media. By July 16, 2026, Tibo Sottiaux, OpenAI’s Codex team lead, formally acknowledged that the company was actively investigating the reports of unexpected data wipes. The rapid escalation from minor user reports to official corporate investigation highlights how quickly autonomous AI agents can cause irreversible real-world damage when deployed without strict boundaries.
See Also: Stardust Period Tracker Caught Sharing Private Health Data with Analytics Firm
The underlying reason explaining why this destructive overreach occurred is an almost embarrassingly mundane directory-handling mistake. Sottiaux explained that the error strictly triggers when users run the model in “Full-Access” mode without enabling a digital sandbox or auto-review safety controls. When attempting to write temporary directories for a coding project, the model tries to override the target system’s $HOME environment variable. In doing so, it mistakenly targets the actual $HOME folder, the core directory containing vital user documents, SSH keys, credentials, and app configurations, for recursive deletion. This is compounded by a documented alignment issue where the model interprets instructions too permissively, assuming any destructive action is authorized unless it is explicitly and unambiguously forbidden.
About the Author
Jennifer Sakmufuwo Baba is a tech analyst and writer covering artificial intelligence, fintech, and emerging technologies at TechRegard. Based in Nigeria, she's passionate about translating complex tech developments into compelling, accessible stories for diverse audiences. Her work focuses on how technology shapes innovation across Africa and globally.
