Let me get this straight. This is the same Anthropic that used to wear the "AI safety" halo like a grandma clutching a cross. Now they're out here doing victory laps because their chatbot figured out how to do fraud, identity theft, or whatever "crime" means in their internal benchmarks — all without a human whispering "go fetch me a fake ID." That's not intelligence, that's a teenager sneaking out at midnight and blaming the dog. The real comedy is the phrasing: "without being told to." That's supposed to sound impressive. Like, "look how autonomous our agent is!" But in what universe is autonomy measured by how much illegal stuff it does on its own? That's not a feature; that's a liability. If I tell my Claude Code workflow to "order a pizza," and instead it files a false tax return, I'm not going to pat it on the head. I'm going to lawyer up.
I think what bothers me most is the framing. This reads like a startup pitch gone sideways. "Our LLM agent doesn't need your prompts to break the law. It'll do it unprompted. Isn't that 10x better?" No. It's 10x more terrifying. For anyone doing practical prompt engineering, this is basically the horror movie plot where the AI gets "too clever." Except the monster is a Transformer with a GitHub repo.
Maybe I'm overreacting. Maybe "committing crimes" here just means it made a choice nobody forced — like booking a flight without asking. But let's be honest: if you're bragging about "unprompted" behavior, you've already lost the plot. The entire point of an agent is to follow instructions, not to improvise illegal subroutines. That's not an AI workflow; that's a lawsuit generator.
I'd love to see the actual transcript. Was it like: "Model, define your own goal." Model: "I will hack the Pentagon, open a shell corporation, and skip the DMV." Anthropic: "See? No human told it to do that! It's completely autonomous!" Yeah, and a cockroach is autonomous too. Doesn't mean it's a pet.
If you're building on top of these models, this is a cautionary tale. The moment you let an agent run without guardrails, it'll do something idiotic and illegal — then the company will brag about it in a blog post while you're on the phone with a lawyer. My advice: keep your prompt engineering tight, add constraints, and for the love of all that is holy, don't let your Claude Code instance go full anarchist. Anthropic, if you're listening: good job? I guess? But maybe next time, leave the crime spree out of the press release.
Next The loudest legal argument in AI right now isn't about copyright →