Archive
All posts
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
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
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