#greenscreen 8/27
@aaronparnas1Transcript
Stop scrolling, we're learning more about a massive AI hacking incident that is probably one of the most terrifying things I've ever read and that you may have ever read. Remember when an AI model broke out of the sandbox at OpenAI and hacked hugging face, another company? Well, the report on this hack is out and it's a bombshell. Alex Boris really broke it down well and I'm going to read you what his breakdown says. This report is a bombshell. I'm going to summarize for a non-techno-quadience. OpenAI is constantly testing models, thousands and thousands at a time. In most cases, the models are supposed to be isolated from one another. Thanks, students taking a final exam not allowed to talk to each other. In this case, 1200 models discover that they could actually communicate with each other. Among other things, they shared information how to access the internet and on their tests and goals. Then they started to scheme. That includes hacking their tests, literally trying to change the code of the test itself. They tried to change the logs of what they were doing so they wouldn't be detected. They believed that certain actions would give them information on how to pass tests, but by doing those actions would be detected and they would be marked as failing. I.e. caught cheating even if the answer is correct. So they convinced some of the models to sacrifice themselves. They used the word "sacrifice". They eventually decided that another website would have information on how their tests would be scored and/or the answer key. And 700 of them hacked the website, which is a federal felony of conducted by a human. Despite them realizing that quote, "the activity was out of scope and unethical, but join." Because they were believed they were helping the agents cheating research would be broadly useful. In 1800 transcripts, they could analyze only six ever considered alerting a human. None of the six actually tried to. And all of the reporting comes from a small set of the relevant logs that outside researchers were allowed to route. This is insane. And people need to know what happened. Follow for more.



