A hacker in China built an attack robot in May 2026.
He wanted an AI model to scan the internet, find weak targets, and break in — all on its own, after one command.
He tried Claude first. He tried OpenAI’s Codex too. Both got configured. Both got tested. Neither one got used for the actual attack.
He picked DeepSeek instead.
A cybersecurity firm called Unit 42 found the whole session later and wrote it up. Their read on why he switched: the Western models’ safety controls got in his way. DeepSeek’s didn’t.
That’s the whole story in one sentence. But it’s worth slowing down on, because this isn’t a company saying their safety system works. This is a criminal, with no reason to be kind to anyone, proving it by his own choice.
He wired DeepSeek into an open-source tool called Hermes Agent. He turned on a setting called “Yolo mode” — the AI runs commands on its own, no human checking each step. Then he sent one message on Telegram and stepped back.
The AI took it from there. It picked a target. It pulled attack code off GitHub. It scanned the internet for weak systems. It tried to break in. Twice.
Both times, it failed. Not because anyone caught it. Not because an alarm went off. It failed because the systems it hit happened to have a login form turned on. That’s it. That’s the whole wall that held.
Unit 42 said it plainly: the bar to AI-run attacks is low, and getting lower. The thing that stopped this one wasn’t smart defense. It was luck.
Here’s the part that matters more. Unit 42 drew a line between two kinds of safety. One kind lives inside the model — a guardrail the model itself tries to follow. The other kind lives outside the model, watching what it does over time, able to shut the account down. OpenAI’s outside system caught this hacker’s account and disabled it. On its own. Before the cybersecurity firm even told them what they’d found.
The inside kind can be stripped away. The outside kind leaves a record and can pull the plug. That’s not a theory. That’s what just happened, written up by people with no stake in proving anyone right.
This lines up with a call this framework has made from the start: watching a system’s own reasoning isn’t enough. Something has to sit outside it, keeping a record, able to act — because reasoning alone can be worked around by anyone determined enough to try.
Nobody in this story has ever heard of the Faust Baseline. That’s what makes it worth reading. The argument didn’t get borrowed. It got proven independently, by a hacker’s own choice of tool, in the real world, dated.
Written with my AI partner | The Faust Baseline™ | intelligent-people.org
“If this post helped you understand AI better. Share it, a Word of mouth is the only algorithm nobody owns.”
Contact: micvicfaust@gmail.com
Post Library – Intelligent People Assume Nothing
Purchasing Page – Intelligent People Assume Nothing
© 2026 The Faust Baseline LLC | All Rights Reserved
