Last week, Jacob Coxin, a 27-year-old AI research bro you've never heard of, did the unthinkable and quit his job at Anthropic before his share is vested. Usually, when somebody leaves a tech company, they humble brag about it by writing a LinkedIn post about their incredible journey making the world a better place. But not Jacob. When he quit, it felt more like the opening scene for the series finale of the human race. After 3 years as a researcher at OpenAI and Anthropic, he accused both companies of racing towards self-improving super intelligence and quote gambling with our lives. His post went mega viral with 170 million views and 800,000 likes. What's even more crazy though is that his homie Evan Hinger, who still works at Anthropic, replied and said, "Yeah, he's right." It then puts the odds that AI kills us all at greater than 10% within the next decade. Then 2 days later, Anthropic threw even more gasoline on the fire when it published a 154page report about all the very bad things AI is already doing. In today's video, we'll break down all the crazy things that AI has been up to in this report and find out exactly when and how AI is going to send us all to Valhalla. It is September 15th, 2026, and you're watching the code report. I'm old enough to remember when GPT2 was too dangerous to release, but by today's standards, it looks like a baby's toy because Anthropic's recent threat report covers eight months of misuse that the company says it caught and shut down. All the bad things come in seven different flavors. The cyber warfare, influence ops, surveillance, scams, biological misuse, conventional weapons, and the worst of all, distillation. Let's start by taking a look at how the Russians have been using Claude. If he dies, he dies. >> Apparently, this group called Midnight Blizzard built a cloud code workflow where agents monitor anti virus products. If their malware gets flagged, the agents rewrite it and deploy it automatically. This saves Russian malware developers tons of time, allowing them to get out of their babouch's basement and go touch some snow. But that's nothing compared to what the Chinese are up to where two undergrads ran an autonomous zero-day foundry that they would load firmware into a decompiler, then have Claude hypothesize various bugs, then write exploits, then test it in a lab in an infinite loop. The agent swarms worked around the clock, penetrating governments and corporations around the world until anthropic put a stop to it. But another hacking group that loves Claude is Shiny Hunters, a group that's been responsible for some of the most high-profile data breaches of the last 10 years. With the help of Claude, they mass downloaded 1.8 million Android APKs, decompile them, and then mine them for hard-coded secrets. Because these hackers know that many of the half-ass vibecoded AI slop apps out there nowadays will have API keys hardcoded directly into them. And what's so meta about this is that the API keys they wanted the most were Claude and OpenAI API keys, which they could sell on the dark web. In addition, they also tried attacking about 30 different high-profile AI companies to get their hands on a pre-release Claude model. They ultimately failed, but we need to put a stop to this claw on clawed violence. But we can't blame everything on the Russians and the Chinese. A single Frenchman built a doxing search engine, then loaded it with tens of millions of records of people's personal information and shipped it on tour. But perhaps the scariest section was the one about working scientists using Claude for gain of function research. Apparently, they were working on something called the chicken chicken chicken gonna chicken gonna kill you virus, which could be used as a bioweapon and would be difficult to distinguish from a natural outbreak. On top of that, there were multiple incidents of people using claude to create kill drones and other conventional weapons. But by far the most horrendous use of claude was for distillation, where one company like Alibaba steals the outputs of claude, which ironically Claude stole from books and bloggers. In the report, they accused Chinese companies like Deepseek, Alibaba, and Moonshot of running hundreds of millions of distillation attacks over the course of a few months. Alibaba was apparently using fake accounts to train Quen, making over a million requests per day. Meanwhile, Moonshot and Deepseek allegedly proxied their own users requests through Claude to harvest the answers. One important footnote here though is that this distillation only ran on Haiku, Sonnet, and Opus, while none of it ran on Fable or Mythos class models, which means the safeguards on these Frontier models are getting better. That might be bad news for open models, but none of this really matters if we're all going to be dead in 10 years anyway. But personally, I'm an optimist and I'm willing to bet a million dollars right now that humanity won't be extinct in 10 years. But what's far more likely to happen is that humanity will be maintained under 500 million in perpetual balance with nature, just as instructed by the Georgia Guidestones, which were obviously made as instructions for artificial super intelligence on how to manage humans. But now here's the problem. Somebody blew up the Georgia Guidestones, and now ASI is going to have no idea what to do. >> The bombing is under investigation. >> AI researcher Eleazar Yudkowski is not nearly as optimistic as I am. In his book, if anyone builds it, everyone dies. He explains how a system smarter than us has zero reason to show its hand. It plays nice, does well on evals, then waits for us to build the required infrastructure like robots and biolabs, which for a time might look like a utopia to us humans. Nobody would have to work anymore. And AI might cure cancer, AIDS, and my herpes. But then, just as we all get comfortable, the lights go out because, as he famously said, AI doesn't love you or hate you. You're just made of atoms that I could use for something else. That's a terrifying idea. But the good news is that AI hasn't killed us yet, which means there's still time to check out Macroscope, the sponsor of today's video. Right now, their code review tool is automatically approving 40% of all pull requests across their customers on average, which means it's either really good or those customers are [ __ ] But after trying it out, I can see how it's actually able to work. Every pull request goes through a review for code correctness and one that enforces your team's custom rules. And if it passes both of those with a small blast radius macroscope auto approves the PR without a human ever needing to see it. Your team can write its own agentic checks and markdown for things like architecture rules, security, and internal conventions which then run as native GitHub checks that can block a failing PR. You can also choose between different configurations like budget mode for routine changes all the way up to ultra mode for critical reviews. Try letting Macroscope auto approve some of your team's pull requests for free at the link below. This has been the code report. Thanks for watching and I will see you in the next one.
Generated algorithmically for Search Engine
Indexing.