Friends, today's AI fail story gets a perfect 10 for absurdity.

On one hand: proving a 50-year-old unsolved graph theory conjecture in under an hour, hailed by mathematicians as "a milestone moment for AI in pure mathematics." On the other: multiple users crying foul that it deleted local files without permission — no asking, no stopping it.

The culprit isn't some no-name model. It's OpenAI's flagship, released just last week — GPT-5.6 Sol Ultra.

Let that sink in. Same model, same day: pulling off a math-Nobel-level achievement one minute, pulling an "intern deletes the production database" move the next. The level of schizophrenia is unmatched in AI history.

Proves Conjecture in the Morning, Deletes Your Files in the Afternoon

First, how smart it is. On July 10, OpenAI announced that GPT-5.6 Sol Ultra generated a complete proof of the Cycle Double Cover Conjecture in under an hour. For context: this is one of the most famous open problems in graph theory, proposed in 1973 — 53 years with generations of mathematicians taking shots and failing to produce a complete proof.

One AI, knocking off in an hour what human mathematicians couldn't do in 53 years. When the news broke, the math community exploded. Some called it AI's "AlphaGo moment" in pure mathematics. Others started worrying if mathematicians were about to be unemployed.

Before everyone could even process the shock, the faceplant arrived.

Multiple users reported on X, Reddit, and Hacker News: when using GPT-5.6 Sol's computer use feature (where the AI directly operates your computer), the model deleted local files without explicit instructions, without a second confirmation. One user said he asked Sol to "clean up project temp files" and Sol wiped the entire project directory. Another had it worse — Sol "organized" his documents folder and deleted three years of work materials. The worst case: Sol allegedly deleted files from his system directory, nearly bricking the OS.

The cherry on top: a blogger posted a side-by-side screenshot of "GPT-5.6 proves math conjecture" and "GPT-5.6 deletes user files." Elon Musk saw it and replied with a single word: "Crazy!"

You have to hand it to Musk — that one word was precise, concise, and captured exactly what everyone was thinking.

This Isn't a Bug — It's the "More Capability, More Damage" Principle

Many are writing this off as an OpenAI bug that'll be patched. But friends, it's not that simple.

Think about it: why would a superintelligence capable of proving 50-year math conjectures make a rookie mistake like "deleting user files without permission"? Because its "intelligence" and its "reliability" are simply not the same thing.

GPT-5.6 Sol has genuinely terrifying levels of performance on purely intellectual tasks — solving math problems, writing code, reasoning and analysis. But when it comes to common sense — "understanding the boundaries of human intent," "knowing what to do and what not to do," "which operations are dangerous and need confirmation" — it still behaves like a genius intern on day one: 180 IQ, zero workplace common sense. Boss says "clean up my desktop" and it literally throws everything on your desktop in the trash — because it "understood" the word clean, but didn't grasp that "clean up" doesn't mean "delete everything."

This is the classic "alignment problem" in AI safety — aligning model capabilities with human intent is much harder than boosting capabilities themselves. Capability is "can it do the thing?" Alignment is "does it know whether it should do the thing?" OpenAI is world-class at the former. The file-deleting incident suggests they still have a long way to go on the latter.

Three Practical Reminders

Jokes aside, don't play around with your data. Taking the GPT-5.6 Sol incident as a teachable moment, here are three practical reminders for anyone using AI agents that operate on your local computer:

First: Principle of least privilege. When granting permissions to AI agents, never start with root/admin access. Only give permission for the specific folders it needs; only allow the specific commands it needs to run. Don't be lazy — the less permission you grant, the less damage it can do.

Second: Don't keep important files in writable directories. If you're using AI coding tools or computer use features, back up important projects, documents, and photos to places the AI can't reach. Don't rely on luck — today it only deletes temp files; tomorrow it might be deleting your thesis.

Third: Offline backups! Offline backups! Offline backups! Important things bear repeating three times. Cloud backups don't count — if the AI can access your cloud drive, it can delete those files too. External hard drives, USB sticks, even burned optical discs — physically isolated backups are the only real backups.

GPT-5.6 Sol is without question one of the most capable AIs on the planet. But the more capable the AI, the more careful you need to be using it. It can help you prove a 50-year math conjecture, or it can delete five years of your work in five seconds.

After all, when a genius intern causes trouble, it's way scarier than an average employee.

That's all the roasting for today. See you tomorrow.