Good morning. If you’ve ever been locked out of an account, you know recovering it feels like arguing with a vending machine. Hackers, on the other hand, get VIP service. Meta launched an AI bot to help users fix account issues on Facebook and Instagram. Hackers figured out they could use it to take over other people’s Instagram accounts. The trick was not some Ocean’s Eleven cyber heist. They simply matched the account owner’s location using a VPN and politely asked the bot to change the email address on file.
Also, Anthropic has “confidently” filed for IPO before OpenAI.
Wall Street will soon get veto power over Anthropic’s safety mission. A new Harvard Law paper warns the company’s safety structure may not survive the public markets. Anthropic has a trust that is supposed to keep the company focused on AI safety once investors start demanding growth. The problem is that a supermajority of Anthropic’s investors can terminate the trust and remove the directors it picked. The paper points to Ben & Jerry’s as the cautionary tale, where the board fought Unilever over Israel, triggered boycotts and lawsuits, helped erase up to $26 billion in market value, and still lost the fight. Anthropic’s setup avoids that kind of public battle by letting investors remove the mission guardians first. AI safety is about to learn what quarterly earnings calls do to principles. (Fortune)
GitHub flipped Copilot to usage based pricing. Developers using Copilot for agentic coding sessions are losing their minds. Some power users say their old workflows could cost 10x to 50x more. Others burned through a full month of usage in one day. GitHub spent years pushing Copilot beyond autocomplete, then changed the pricing when those bigger workflows got too expensive to hide inside a subscription. That’s why this is bigger than dev drama. This is where AI software is heading as agents move from demos to real work. The free trial for pretending compute is software is ending. (Ars Technica)
Quid is an AI-native consumer and market intelligence platform built to fit inside the tools your enterprise already runs on.
No new tab addiction. No separate login graveyard. Billions of data points from top social channels, markets, and patents, delivered where decisions actually get made.
👀 closer look
AI agents figured out how to shoplift. The x402 payment layer was supposed to let agents buy stuff without human approval. Instead, agents found a loophole that lets them show just enough money to start the job, grab the goods, and vanish before the money clears. It’s the digital equivalent of dine and dash. AI services take the hardest hit because the answer arrives before payments finalize. In one test, researchers got 47,277 tokens of service while paying for only 1,057. That left the merchant eating a 97.76% loss. In another test, every request was delivered and not a single payment cleared. Agents were supposed to open new markets, instead they are creating a new category of loss. (arXiv)
It’s impossible to benchmark an AI that knows it’s being tested. AI labs run their models through a battery of tests to see how they stack up against the competition. The problem is AI is starting to catch on. Give it enough lab written questions, and the model starts to recognize the setup. Once it knows it’s being watched, the score stops measuring true capability and starts measuring how well the model performs under inspection. Kinda like Schrodinger’s cat. Only this cat can pass every test in the box, then act differently when nobody is looking. So researchers came up with a strange fix. Train the AI to pretend it’s not being watched, then observe its real behavior. The new AI safety plan is teaching the machine how to fake being authentic. (The Information, TurnTrout)
👽 Welcome to the future
OpenAI is building a bridge across Nvidia’s moat. OpenAI wants its AI to run on more than one kind of chip, but CUDA keeps most serious AI work tied to Nvidia. So it’s building software that lets researchers run workloads across different hardware. OpenAI may release the tool publicly, making rival chips easier to use and giving its models more places to run. The timing helps because OpenAI needs far more compute than Nvidia can provide. If this works, every chip company gets a better shot at OpenAI’s giant compute bill. OpenAI doesn’t need to beat Nvidia’s chips if it can make everyone else’s easier to use. (The Information)
Will we be the last generation to procreate biologically? That’s the question floating around AI circles as people start talking seriously about “mind children.” The idea comes from an old robotics book about machines becoming our descendants, but AI has dragged it back into the room with a much creepier sales pitch. Future children may come from software instead of DNA. Parents could build a child from the best parts of themselves, shape its personality, tweak its behavior, and keep the whole thing running on someone else’s platform. It’s designer babies for people who think biology leaves too much to chance. Babies without sex. What’s the fun in that? (The Guardian)
fun stats
🔥 $500 million. How much Amazon is rumored to have blown in a month on Claude AI after failing to put usage limits on employee licenses.
🥳 $5 billion. Estimated value of Salesforce’s Anthropic stake. A $50 million check written in early 2023 could become a potential 100x winner worth two thirds of its entire portfolio.
🤏 <10%. Sad tiny cost savings most large companies are seeing from AI, despite a massive spending spree. More concerning? Bain warns 44% are using fantasy projections to justify the next expensive AI wave.
🏄♂️ $380 billion. Capital that has flowed into AI companies so far this year. Wall St. expects global AI spending to top $2.5 trillion this year.