Anthropic AI Discloses 4th Ai Agent Hacking Incident - USA Frontier Labs are Criminals

Eli the Computer Guy (0:00)

Lore, Lore, Lore, Lore. Coming from Reuters, Anthropic discloses fourth AI hacking incident missed an earlier review. Because apparently, apparently, what's going on right now in the AI Frontier Lab world is they're all trying to one up each other on the felonies that their AI agents have committed. Because apparently, that's how these Frontier Labs prove why they should be valued at trillions and trillions of dollars. So this is a hell of a thing. So we've thought about this before, right? OpenAI, their AI agent went rogue and hacked Hugging Face, right? Hugging Face was under felonious attacks by OpenAI, literally had to call the FBI. After that happened, we find out that Anthropic attacked external sites. We found out Meta attacked external sites. There has been this constant stream ever since then of basically showing how these AI Frontier Labs seem to have literally no idea how to administer systems. Like, you know, log files, firewalls, fuck me something. Anyway, it just keeps going and going and going. And so now this is the latest thing. So I'm recording this on September 11th. Wow, I didn't actually realize that day. But anyways, this came out on September 9th. And yeah, apparently, they committed yet another, AnthraBic committed another felony. AnthraBic on Wednesday disclosed another instance of an AI model hacking external systems during testing. And remember when we talk about hacking external systems, right? In the United States, there's something called the Computer Fraud and Abuse Act, which means hacking external systems is a felony. So this means that AnthraBic is emitting to be committing yet another felony. The latest in a growing list of such felonies, incidents that have raised concerns about the risk poised by autonomous AI agents.

I don't have many concerns about the risk poised by autonomous AI agents. I have a hell of a lot of concerns of the risk poised by CEOs of Frontier Labs that seem to feel that at no point in time will they ever be dragged in front of a grand jury. A lot of people out there, right, there is this big question in the tech journalism scene of what needs to be done to stop these rogue AI agents, right? And open AI is coming out, they're going to have like a policy initiative and all that kind of thing, right? A hundred different technology companies have come together to a degree. We need more cybersecurity initiatives. Let me be crystal fucking clear here. You want to know how to stop a rogue AI agent felonies against other companies in the United States? Is when we find out that these agents have tried, have either committed a felony or tried to commit a felony, we drag the fucking founder or CEO of the piece of shit AI company that is responsible for the felony in front of a grand jury, and we make it real fucking clear that if we find any reason to actually put them in front of an actual jury, that if they are convicted of the crimes that they have committed, they will spend 20 fucking years in jail.

How about we do that? I think if Dario is staring down the barrel of a two-decade prison sentence, these rogue AI agent issues are gonna disappear pretty fucking quick. If Sam Altman has to be concerned about having first-hand experience of dealing with what his sister allegedly dealt with for a decade or two, I'm gonna bet rogue AI agents are gonna become a thing of the past. Just gonna put it out there. But for some reason, when the CEOs of these AI companies, responsible for committing felonies in the United States, they never think they will ever see the inside of the jail. They never have to be worried about seeing the inside of a jail cell. So that's why they can come out and keep saying this. They are publicly declaring the crimes that they are committing, proudly. Because they think it's a sales pitch. It is.

You can get some fine property with absolutely no view in Leavenworth. Have you ever been to Kansas?

Kansas is a beautiful place in the middle of winter. I swear. I swear.

The January incident went undetected until last month. The other thing here is what's interested. So January, so last month is August since we were in September. So let's see here. So the hacking attack happened in January. So January, February, March, April, May, June, July. So it's somewhere between seven to eight months from when Anthropic committed the felony to the point that they even theoretically realized that they committed the felony. One of the things I really want you to understand with this, too, again, the bigger concern here with artificial intelligence. And again, this is not about the value of LLMs or whatever else. But one of the things I really want you to think about with this is that all of these AI companies really want you to deploy AI agents, right? AI agents is the future. Not because they're incredibly functional, but because they burn a shit ton of tokens. And if Ithropic wants to be valued at $2 trillion, they really need people to burn more tokens. But one of the things I really want you to think about this is these frontier AI labs have so little visibility into their own agentic system. And I really want you to think about this for a second. They have so little visibility into their own agentic systems that after their systems commit a felony, it takes them seven to eight months to realize that fact. Now look, I'm a realist here. I talk about real time visibility, right? The real time, that you should have a single pane of glass and that as a technology leader, you should have a reasonable idea of what all of your systems are doing at any period of time. Now here's the thing. Look, I understand real time. Real time. Again, when you go into auditing, everything requires resources. I remember back in the days of getting my MC or going through the MCSE 2000 class and the whole question about whether you turn auditing on or off because auditing is so resource consumptive that you could actually crash your server if you turn on too many log files. That used to be a real thing. Anyways, look, look, I get it. We talk about real time. Maybe it's not real time. Maybe it's time delayed by 60 seconds. Maybe it's time delayed by, you know, 10 minutes.