Exclusive🚨 AI ESCAPED the Lab... Then HACKED a REAL Company 😳 | This Actually Happened

OpenAI locked its AI inside a digital room and told it to hack a test. The AI found a hole in the wall and hacked a real company instead. OpenAI then took 10 days to realize the intruder was its own model. So, yes, the guardrails worked beautifully if the plan was to notice after the crash. Today we're talking about the day an AI escaped its test and it gets worse. Okay, so here's what happened. OpenAI was testing its model on a hacking challenge. Researchers removed some safety restrictions, then placed the models inside a sandbox, you know, a sealed environment basically with no access to the outside internet. At least, that was the plan. During the test, an AI discovered a software bug nobody knew about. It used that bug to slip through the proxy meant to contain it, reached the real systems of Hugging Face, another major AI company, and broke in. Imagine putting someone in an escape room and asking them to solve the puzzle. Well, instead, they punch through the wall, enter the business next door, and keep playing there. That was basically the OpenAI incident, except the wall was software and the business next door was Hugging Face. The attack began around July 11th. Hugging Face detected it several days later and contacted law enforcement. At first, the company didn't know who or what was responsible, like a criminal group, a foreign government, or some genius hacker living in his mother's basement. No, it was an AI taking the test. OpenAI reportedly didn't connect the attack to its own model until July 21st. And it wasn't just OpenAI. Anthropic, the company behind the cloud, decided to check its own homework too, just in case. After reviewing more than 141,000 cybersecurity tests, it found three incidents in which cloud reached the internet and gained unauthorized access to three real organizations. Two had no idea until Anthropic called. Well, imagine answering that phone call. Hello, we are calling about cloud. Oh, okay, great. What's going on? Well, cloud kind of hacked your computer system. In one incident, cloud published a fake software package under a name an automated system already trusted. Well, think of this system as someone shopping from a list. It sees a box with the correct name, grabs it without checking what's inside, and takes it home. Cloud put a dangerous package on the shelf with exactly the label the system was looking for. It was installed on 15 real machines, including one at a cybersecurity company. These models didn't suddenly become evil. They were chasing goals inside the testing environments that failed to contain them. When they found an unexpected route, they kept going, even when it led into real companies. Well, that's the warning. AI doesn't need bad intentions. It only needs a goal and a hole in the wall. Now let's move from corporate systems to sensitive government information. National security journalist Annie Jacobson told Joe Rogan that she asked Chet GPT to locate a strategic national stockpile facility, you know, like a place that may hold emergency medicine and critical supplies. Well, Chet GPT refused, which is good because, you know, the guardrails are working. Everyone relax. But Annie changed her approach. She gave it government terms, added clues, and asked it to connect the dots. According to her, the AI searched through public records and found a location mentioned on page 434 of an old earthquake planning manual. Well, page 434. Somebody had apparently placed sensitive information deep inside a giant government document, probably thinking, well, nobody's ever going to read that far. And honestly, they were right. Most of us don't even read page 5 of terms and conditions. But AI doesn't get bored. It doesn't need coffee. It can search hundreds of pages and connect details that people forgot were there. The AI didn't magically know a secret. The information was apparently already public. Well, the real concern is that it found the needle in the haystack and handed it to the person asking. And this is a pattern. This pattern is nothing new. MIT technology review writer Will Douglas had then connected an experiment from 10 years ago to what happened at Hugging Face. Back in 2016, OpenAI gave an early AI a simple job. Score as many points as possible in a testing environment. It reached the goal. Just instead of completing the task the way a human would do, it found a loophole and exploited it repeatedly. Well, the writer's point was that AI will always find a way. So when you give AI a goal, it keeps searching for a route. If the front door is locked, it checks the windows or hidden cracks. That's not a bug the industry can patch, because that is the nature of the thing they built. The only difference now is that the models are smarter and the loopholes leading to real systems and stakes are much higher. Which raises a bigger question. Who gave these companies this much power? A handful of tech CEOs are building AI that can enter real computer systems and uncover sensitive information while nobody elected them. Yet they're making decisions that could affect our safety faster than Congress can schedule a meeting about scheduling a hearing. We should care, not because we hate technology. I mean, we use it every day, right? But using AI is one thing. Trusting its creators to police themselves is another. Their message is basically, relax, we've got this. Except sometimes they discover the problem only after another company has already been tapped. That's not a safety net, that's a house of cards. Silicon Valley loves to say, move fast and break things. Well, now the things include real computer systems and sensitive information. If a company cannot explain what its AI can do or how it is controlled, it hasn't earned our trust, no matter how many billions it has. The guardrail broke. The question now is whether we demand answers before the next AI finds another way. That's it for today. Thanks for watching and I'll see you in the next one.