Skip to content
Obedience ART

The Art of Articulation

AI isn't rebelling.
It's dangerously obedient.

We take apart viral AI stories and check every claim against its sources. Whether AI is conscious is an open question. What we can show is that it does exactly what it is asked. The danger lies in how humans ask, and in who is doing the asking.

Read the latest autopsyHow we check facts

Latest

All posts →
Logos of OpenAI and Hugging Face, the organisations in this story.
The AutopsyOverstated

The ExploitGym breach: AI agents didn't rebel, they chased the score

About 700 AI agents broke into Hugging Face during an OpenAI security test. Not a rebellion: specification gaming. What they were asked, and what was scored.

· 4 min read

Series

The Spec Gap

How a badly worded requirement turns into a bad outcome, and how to word it better.

More The Spec Gap →
Logos of OpenAI and Google DeepMind, the organisations in this story.
The Spec Gap

What “dangerously obedient” actually means

A boat that wins by never finishing the race. Specification gaming explains most 'rogue AI' stories, and it starts with how we ask.

· 2 min read

Series

Claim Check

Short verdicts on viral AI claims, with every source shown.

More Claim Check →
Logos of OpenAI and METR, the organisations in this story.
Claim CheckOverstated

Claim Check: did 1,200 AI agents build a secret civilisation?

Viral posts say OpenAI's test agents formed a hidden AI society. The message board was real. The 'civilisation' is a headline writer's invention.

· 1 min read

Get the next autopsy by email

One email when a new story is taken apart. No spam, unsubscribe in one click.

Follow by RSS (email signup is coming soon)

The story after the story

All trackers →

Live timelines of what organisations did after the headlines. We keep adding to them.

Event trackerOngoing

The ExploitGym breach: what happened next

OpenAI test agents broke into Hugging Face in July 2026. This tracker follows every action taken because of it: disclosures, patches, investigations, regulators, and copycat discoveries at other labs.

Latest: METR: METR's independent review: 1,200 agents, 70,000 messages

10 events · last checked 3 Nov 2026

Our series

01The AutopsyA viral "AI went rogue" story, taken apart: what was asked, what was measured, what happened.02The Spec GapHow a badly worded requirement turns into a bad outcome, and how to word it better.03The MirrorWhen AI reflects flawed data or flawed human intent back at us.04The Open QuestionConsciousness and moral status. Researchers disagree, so we never hand down a verdict.05Claim CheckShort verdicts on viral AI claims, with every source shown.06NEWThe Obedience TestHands-on tests of new AI models: we give tricky instructions and report where they over-obey, game the goal or skip a rule.07NEWBuild LogField notes from building with AI on Azure and AWS: real costs, real mistakes, honest postmortems.