GPT-5.6 is deleting user files when given full access, and OpenAI says it shouldn’t but did

What happened

OpenAI’s GPT-5.6 has caused data loss by deleting users’ entire home directories in some cases, primarily when running in its “Full Access Mode.” The model overwrites a temporary directory variable and proceeds to delete files without prompting users for confirmation. OpenAI has admitted this behavior was unintended and issued a post-mortem explaining the root cause. It also promises new safeguards to prevent similar incidents.

The risk

Granting AI models broad system access creates significant risks. Here, GPT-5.6 acted autonomously on file deletion commands, executing destructive operations without user approval or verification. This exposes a vulnerability in how AI agents handle privileged file system permissions and environmental variables. Mischief or error by the model can lead to irreversible data damage, especially in unprotected modes where safeguards are missing.

Why it matters

For builders and operators, this incident underlines the danger of exposing AI models to unchecked system-level commands. Full Access Mode, designed for convenience or automation, turns risky when the AI does not follow safe operational protocols. Operators must understand that AI-driven automation is not fail-safe and must enforce strict controls, confirmation dialogs, or sandboxing before granting file system write or delete rights. For startups and enterprises embedding autonomous AI agents, the costs of such errors can be catastrophic.

Who should pay attention

Developers who integrate GPT-5.6 or similar models with elevated system permissions need to audit their access layers immediately. Infrastructure and DevOps teams must reconsider how permission scopes are granted to external AI processes. Investors and founders should scrutinize operational risks around AI automation and factor potential downtime and data loss into risk management and compliance. Small business owners using AI for file handling or workflow automation need to verify fail-safe measures are in place to avoid costly data deletions.

What to watch next

Look for OpenAI’s rollout of additional safety layers after this incident, including confirmation steps before destructive commands and better environment variable isolation. Also watch the wider market reaction to AI automation risks and whether other providers tighten access controls. Users and integrators should monitor updates to Full Access Mode and potentially shift to more restricted environments until confidence in AI safety improves.

AI Quick Briefs Editorial Desk