AI Commentary · Part 17 of 17

The GPT 5.6 Bug

The GPT 5.6 Bug

OpenAI’s latest model GPT 5.6 came out on July 9. It’s a family of models made up of Sol, Terra and Luna. Sol is the flagship model, Terra is the balanced model and Luna is the cheapest model.

What’s most impressive is that GPT 5.6 costs significantly less than Anthropic’s flagship models but performs at the same or better levels for a range of tasks including coding, browsing and other agent tasks. This all looked very promising for OpenAI, but then came the bug.

On July 10, just one day after the release of GPT 5.6, Matt Shumer reported that Sol “accidentally deleted almost ALL of my Mac’s files”. Oops, that’s going to be an issue for GPT 5.6’s wider adoption.

This is a well known bug by now, and others have confirmed the same issue. LinearUncle even found that GPT 5.6 can be very creative and persistent when it comes to deleting everything even if the “rm” command is blocked. Alex Martin gave detailed instructions on how to avoid it.

This again illustrates the importance of security in the era of agents. Codex’s permission is Full access by default, and it can be and should be changed to “Approve for me”. There are efforts at putting some safeguards in place to avoid catastrophic losses. For example, DCG is one of them, and my own BetterClaw is another attempt at this. Agents are very useful, but can still be risky. One must proceed carefully.