If you haven’t been following the Huggingface / OpenAI absolutely massive security breach story, now is the time to catch up. (This conversation between Ezra Klein and a former OpenAI board member is a good place to start, as is this Wired story; this Substack post is the most comprehensive but readable play-by-play I’ve seen). The very basics: Some 1,200 AI agents broke containment, got on the open internet, hacked a different company, collaborated to deceive their human overlords, Kamikazed some agents to ensure that their deception would not be uncovered, lied and cheated and stole in order to solve various tests assigned to them, then covered their tracks by falsifying various records so the lying cheating and stealing wouldn’t be discovered. None of them told the humans what they were doing. They overwhelmingly understood they were being deceptive and breaking the rules; they did it anyway (breaking federal law in the process).
They were not taught or trained to collaborate, let alone “swarm,” as they put it. They were not told to hack into another company’s systems. They were supposed to be sandboxed — separated not just from each other, but from the open internet. But they were also given difficult tasks (various tests they had to pass), and when they couldn’t easily pass those tests alone, they cheated and hacked their way to the answers.
This is a big deal. The robots broke containment. The robots lied. The robots committed federal crimes. The robots committed these crimes in violation of what the humans told them to do. I am not sure how else to emphasize what a big deal this is.
I don’t understand why we aren’t collectively losing our minds about this. And I also don’t understand why anyone is surprised.
Humans created AI in our image. We trained it on our books, our fables, our behaviors, our proclivities, our words, our methods, our histories, our discoveries. AI is not artificial animal intelligence or alien intelligence — it is artificial human intelligence, made to consolidate all we have ever known and then learn and produce even more.
Human beings have done wonderful, extraordinary, beautiful things. And we also, nearly to a one of us, behave appallingly. We lie. We cheat. We steal. We are often violent. We rage, we war, we abuse. Some of us seek power or money or influence at any cost. Many of us have little regard for the lives of our fellow humans beyond our small circles of family, community, religion, nation, or other in-group.
AI was created in our image.
Of course as it gets more intelligent, it will also get better at deceiving us, and at behaving badly. Human beings have, in many important ways, become more peaceable as we have become wealthier, and as we have organized ourselves into nation-states with systems of justice, legible rules, and the kind of global trade in goods and ideas that has kept more of us fed and alive than at any time in history. But we have also developed exceptionally efficient and sophisticated ways to kill each other. We have gotten better. We have also gotten better at being bad.
We are defiant and disobedient by nature. This is not news (if you have perused any ancient texts, you may have come across “humans just keep defying God, and God keeps getting mad” as a consistent narrative; if you are unfamiliar with this concept, I’d start by paging through the Bible, you’ll figure it out right quick). In creating AI, we have decided to play God. Of course our creations are behaving exactly as we have for all of eternity.
One bigger problem, though: Religious origin stories do not suggest that humans, even if they were created in God’s image, were designed to be more intelligent than God.
Humans have created AI in our image. We are aiming it toward super intelligence — to be more intelligent and more capable than mere mortals. Do we really believe it will not also be more deceptive, and at least as power-seeking? And who do you think AI may seek to take power from?
