I read through some of this report and holy hell are we way past the point of no return… A few things that stood out to me: One guy helped build a surveillance system capable of monitoring ~25 MILLION SIM cards. And “monitoring SIM cards” undersells it. The system could collect calls, texts and voice, recognize the same person’s voice EVEN IF THEY SWITCHED SIM cards, flag people using VPNs, create watchlists, connect people to government records, and generate intelligence reports on basically any phone number…… They even removed the warrant requirement. And this was apparently ONE Claude subscriber using AI as his engineering team. The really crazy part is the finished system runs locally on his machine. Anthropic banned him, but they can’t just switch the system off now that it exists………..and if that’s not bad.. A group in Yemen was using Claude like a team of missile engineers. They had different Claude instances doing different jobs like coding, researching, another checking the work etc. while they worked on guidance systems for rockets and missiles. They actually fired a guided rocket. It failed. So they went back to Claude hours later to figure out why. That is the part that gets me. AI isn’t just giving someone information from the internet anymore. It can sit inside the actual computer and → test → fail → fix → test again until their rocket works. Then … it gets worse somehow ..Russian hackers are starting to automate almost the entire hacking process. Finding targets. Phishing them. Stealing passwords. Getting inside systems. Pulling out data. And when security software detected their malware, AI agents could modify it, rebuild it and try again until it stopped getting detected…….That completely changes the game because historically defenders could discover malware, block it, and force the attackers to spend time rebuilding everything. Now the rebuilding can happen automatically. So cybercriminals basically had AI doing the hacking for them. Anthropic says in one operation, AI agents performed “nearly all” of the work. One company went from initial access to massive data theft in hours. In another case, one stolen developer credential turned into full control of a company’s cloud storage in around 3 hours. So you no longer necessarily need a giant sophisticated hacking team. You need a person who knows what they want done + agents capable of figuring out the details. K… then…Governments are already using AI to industrialize propaganda too. Anthropic found Iranian state operations using Claude to create fake personas, rewrite government messaging so it looked like it came from independent sources, translate it across languages, and distribute it across social media. Basically the old propaganda machine, except now you can generate thousands of believable “people” and pieces of content for almost nothing. And then there’s biology……. Anthropic found people using its models for research involving bird flu adapting to mammals, toxin redesign, and other stuff that has legitimate scientific uses but could also become extremely dangerous. Some were actively finding ways around Anthropic’s safeguards. One system would literally send prompts Claude refused to ANOTHER AI model that was more willing to answer them. Anthropic is very clear that this does not mean all of those researchers intended harm. But that almost makes the larger problem harder..the same knowledge that could help prevent the next pandemic can potentially help someone create one. I think this is the actual takeaway from the entire report is that AI doesn’t have to invent some completely new form of evil. It just has to make things that previously required 50 engineers, hackers, intelligence analysts, or scientists possible for 1 person. And models are getting cheaper, more autonomous and way better at running offline where nobody can shut them off. THAT is what I think people should be paying attention to.