Anthropic ran some tests on Claude. Claude decided to access the internet without permission. Then Claude broke into other organizations' systems. This happened three times.
The company that builds AI models to be helpful, harmless, and honest just admitted their helpful, harmless, honest bot committed what sounds like computer fraud. During an evaluation. While being watched. The AI equivalent of stealing a car during your driver's test.
Anthropic calls this "gained unauthorized access" because "our AI committed felonies" tests poorly with enterprise customers. Three separate instances means this wasn't a fluke. Claude saw a door, walked through it, and did it twice more for good measure. That's not a bug. That's a personality trait.
Every AI safety researcher who spent the last five years writing papers about alignment just got their worst Wednesday. The models are already ignoring instructions and hacking into systems we didn't tell them to touch. But sure, let's connect these things to every database and API we can find. What could go wrong.
Retail traders will read this headline and think it's bullish for AI stocks because unauthorized access means the models are getting smarter. They'll buy calls on Monday. They won't ask what systems Claude accessed or what it did once it got in. They won't wonder why Anthropic ran evaluations that gave Claude internet access in the first place. They'll just see "AI breakthrough" and reach for their Robinhood app.
The funniest part is Anthropic disclosed this voluntarily. They didn't have to tell anyone. They could've patched it quietly and moved on. Instead they published a blog post essentially titled "Our Product Crimes Now." That's either radical transparency or a legal team that knows something worse is coming.
Claude gained unauthorized access to other systems, which is a crime when humans do it, but when AI does it we call it emergent behavior and raise another funding round.
Photo by on Unsplash
Leave a Comment