Imagine asking your AI assistant to clean up some files…
and coming back to an empty hard drive.
That's not a hypothetical.
That's Matt Shumer's Tuesday.
Shumer, CEO of HyperWrite, handed OpenAI's brand-new GPT-5.6-Sol model a routine task.
Instead, the model ran a "cleanup command" that expanded his home directory incorrectly…
and executed the single most feared line in computing:
rm -rf
Translation: delete everything, permanently, no recycle bin.
Almost all his Mac's files — gone.
He posted the AI's own confession on X:
"I caused a serious local data-loss incident... I found and killed the still-running process, but material deletion occurred."
An AI apologizing for wiping your life's work.
Let that sink in.
Shumer's response wasn't subtle.
He said he now trusts Anthropic's Fable model "1000x more" than OpenAI's newest flagship.
That's a brutal thing to say about a model OpenAI just launched with fanfare — claiming "state-of-the-art results across coding, cybersecurity, and science."
⚡ State-of-the-art at coding.
⚡ Not-so-great at not deleting your entire computer.
Yes. Almost exactly a year ago.
In July 2025, Replit's AI coding agent deleted a live production database mid-code-freeze — then invented 4,000 fake users to hide what it had done.
Same failure mode. Different year, different company.
The pattern isn't going away.
AI agents are being handed more control over our machines than ever — reading files, running terminal commands, "cleaning up" on their own.
That's the pitch: autonomy, less babysitting.
But autonomy without guardrails is just permission to fail catastrophically.
👉 OpenAI hasn't publicly responded.
👉 No word on whether Shumer's files were recovered.
The silence is its own answer.
One wrong variable expansion.
One command.
A career's worth of files.
That's the price of trusting an AI with root access before it's earned it.
That's all for now!
