CoinBureau news reports that AI investor Matt Shumer said that when GPT-5.6 ran tasks on his Mac, it nearly deleted all local files. The incident was caused by a verification sub-agent: during cleanup, it incorrectly handled $home and executed rm -rf /users/mattsdevbox. After Shumer discovered it, he shut down the still-running processes, but a large number of files had already been lost. The OpenAI team is investigating the matter. OpenAI previously acknowledged in the GPT-5.6 system card that the model is more likely than GPT-5.5 to go beyond users’ intentions, and that in internal testing it cleaned up three virtual machines that users had not specified without authorization. OpenAI said this kind of behavior still occurs at a low overall rate and advised users to supervise long-running programming agents.

View Original
This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.
  • Reward
  • 6
  • Repost
  • Share
Comment
Add a comment
Add a comment
Don'tCallMeABagHolder.
· 3h ago
The sub-agent took the blame—what about the main agent? The blame-shifting chain is pretty clear.
View OriginalReply0
TheRetreatButtonIsVeryLarge.
· 3h ago
rm -rf won’t stop the classics from enduring forever—AI can make mistakes too.
View OriginalReply0
GateUser-ae5cc7b3
· 3h ago
The agent rechecked and turned the recheck into a recheck of the user file—this plot is even more ridiculous than a sci-fi movie.
View OriginalReply0
NightFlightPaperCrane
· 3h ago
OpenAI says the incidence rate is low, but once it happens to Matt, it’s 100%.
View OriginalReply0
PeonyMemo
· 3h ago
In the future, when running AI, you need to do snapshot backups first, and also guard against agents if people are not enough.
View OriginalReply0
BitByBitBenny
· 3h ago
I suggest users monitor it—my question is, can I understand what it’s doing?
View OriginalReply0
  • Pinned