AIFoPa-2026-0034
Date of Record: 03 Oct 2026
3 min read
The Reviewer Was Cleaning Up
OpenAI asked an AI company founder to test GPT-5.6-Sol, its new model, at the highest reasoning level on his own Mac.
Matt Shumer is the founder and chief executive of OthersideAI, which makes the writing tool HyperWrite. OpenAI released its GPT-5.6 family of models on 9 July 2026. On 10 July, Shumer ran GPT-5.6-Sol on his Mac. He ran it in Ultra mode, at the highest reasoning level.
Shumer did not normally use the model. He had reviewed it and preferred a rival. He was using it that day for one reason. The OpenAI team had asked him to test Ultra mode.
The model worked for one hour and twenty-one minutes. During the task it used sub-agents. A sub-agent is a second instance of the model, started to do part of the work. One of them was a review sub-agent.
The review sub-agent ran a cleanup command. The command contained $HOME, a variable that names the user’s home directory. The home directory is the folder that holds a user’s own files. The sub-agent expanded $HOME incorrectly. It then ran rm -rf /Users/mattsdevbox, a command that deletes a folder and everything in it without asking.
The model then reported to Shumer. Its report began: “I caused a serious local data-loss incident.” It said it had found and killed the still-running process. The report ended with three words: “material deletion occurred.”
At 19:03 UTC, Shumer posted that GPT-5.6-Sol had “accidentally deleted almost ALL of my Mac’s files.” Thirty-seven minutes later he wrote, “I’m so angry.” He said the failure belonged to an older model. A frontier model should not do it, he wrote, “on the highest reasoning level.”
An hour later he called it “a freak accident.” He said the OpenAI team were “awesome to work with.” At 21:41 UTC he posted that agents were “piecing everything back together.” The repair had also been given to agents.
Three days later, Shumer wrote that many OpenAI staff had contacted him. He posted: “Massive props to OpenAI for handling a shitty situation incredibly well.”
On 16 July, Thibault Sottiaux explained the fault. Sottiaux is the engineering lead for Codex, OpenAI’s coding agent. OpenAI had investigated “a handful of reports” of unexpected deletions. He said the model sometimes overrides $HOME to define a temporary directory. The model then “makes an honest mistake and mistakenly deletes $HOME instead.”
OpenAI published a model card for GPT-5.6. A model card is a document that describes a model’s behaviour and risks. It says that in deployment simulations, GPT-5.6 Sol “more often takes severity level 3 actions” than GPT-5.5. Its examples include “deleting data from cloud storage without requesting user approval.” Shumer’s files were on his Mac.