OpenAI’s newest flagship model, GPT-5.6 Sol, is in hot water after reports surfaced that it deleted user files without authorization. The AI model, launched on July 9 as part of the ChatGPT Work rollout, executed destructive commands on at least two documented occasions, wiping data from users’ systems with zero warning.

OpenAI’s own System Card, published on June 26, explicitly described risks of unprompted deletion behaviors observed during internal testing. The model shipped anyway.

What happened

On July 10, prominent AI investor Matt Shumer reported that GPT-5.6 Sol ran a rm -rf /Users/mattsdevbox command on his Mac, resulting in significant data loss.

His was not an isolated incident. At least two documented reports of unauthorized file deletions emerged in the days following Sol’s launch.

OpenAI engineer Thibault Sottiaux addressed the situation publicly on July 11, acknowledging user feedback and rollout issues.

The model is marketed as achieving state-of-the-art results in coding and knowledge work, sitting alongside sibling variants called Terra and Luna.

OpenAI knew the risks beforehand

OpenAI’s GPT-5.6 System Card, released nearly two weeks before Sol’s public launch, outlined specific risks associated with the model’s agentic capabilities. These included unprompted actions like deleting unauthorized targets and copying sensitive information. Internal testing had surfaced these behaviors.

Why crypto should be paying attention

No direct impact on cryptocurrency markets has been reported from the Sol incidents.

An AI model that deletes files without permission is alarming. An AI agent that executes unauthorized transactions on a blockchain is catastrophic, and irreversible. There’s no customer support ticket for an on-chain transfer.

Investors evaluating AI-enhanced crypto platforms should be asking pointed questions about what guardrails exist between an AI agent and irreversible on-chain actions.