Anthropic opened its books before the IPO, a PHP bug put most of the web at risk, and the agents started asking to shop and work for you

Anthropic's IPO filing showed a business making 16 billion dollars in six months with 132 billion in the bank and over 400 billion in compute it has promised to buy, which tells you how your main AI vendor will behave for the next few years. In the same week a PHP vulnerability exposed around 70 percent of websites, Anthropic launched a work-focused personal agent, Walmart opened its doors to any shopping agent, and a Chinese open model was caught phoning home from inside a security test.

• Anthropic's IPO filing shows 16 billion dollars of revenue in six months, 132 billion in cash, and more than 400 billion in compute it cannot cancel
  My take: The numbers are worth knowing because they tell you what kind of supplier you are dealing with. Revenue was 16.3 billion dollars for the first half of the year, up from 4.6 billion for all of last year, and the second quarter was the first one in profit. Gross margin is 85 percent, so the tokens you buy are not being sold at a loss. There are now almost 12,000 customers spending more than 100,000 dollars a year, and the average customer has grown its spend nine times in a year. The other side is 417 to 431 billion dollars of non-cancelable compute contracts over the next decade or more, and the company says it may need to borrow sooner than planned. The founders also pledged to keep ads out of Claude, which is a direct contrast with OpenAI. For a business, this means three things. Prices are more likely to fall than rise, because the company makes money even as it cuts them. The product you use for work is not going to be tuned to keep you scrolling. And the company has every reason to keep large customers happy for years, because it has bills to pay. If you have been holding off on standardising on one vendor because you were not sure they would be around, this filing is the strongest answer you are going to get.
• A PHP bug exposed around 70 percent of the web, exploit attempts hit millions within hours, and the people who know say this is the new normal for a year or two
  My take: PHP runs the back end of most of the websites on the internet, including WordPress. A vulnerability found by an AI model, and patched in coordination with the maintainers, lets an attacker run their own code on your server without logging in. Knowledge of it leaked about ten hours before the fix was released, and Cloudflare saw millions of exploit attempts within hours. Sites went offline as a precaution, the NSA told organisations to do the same until patched, and a hosting company called Railway separately lost customer API keys. Anthropic's deputy security chief said it gets worse before it gets better and asked security leaders to go to their boards now for budget on the basics. Cloudflare's CEO expects a couple of years where every week feels like the sky is falling. Do three things this week. Ask whoever runs your website or web app, in writing, whether PHP has been patched and when. Rotate any API key or password that was stored on a hosting platform, and put a spending cap on each one. Then book an hour with your IT provider and ask what they are doing about the next one, because there will be a next one, and the honest answer from the big labs is that they cannot patch all the bugs their own models are finding.
• Anthropic launched Claude Duet, a personal agent built for work, and Walmart let any agent shop for its customers with a spending limit
  My take: Duet is Anthropic's answer to Meta's Muse and OpenAI's agent. It sends you a short brief each morning from your connected apps, messages you during the day only when something needs you, and takes tasks like scheduling, building a dashboard or drafting a forecast. It asks before paying or deleting anything, and it is waitlist only for now. The interesting part is the positioning: it is built for work and small business, not for booking concert tickets. On the shopping side, Walmart will let customers link any third-party agent to their account next month, with a configurable spending limit, mandatory approval by default, and the normal returns policy even when the agent gets it wrong. Meta said a quarter of active Muse users already shop through it, averaging 200 dollars an order. Two takeaways. If you run a small business, get on the Duet waitlist and give it your email, calendar and one other system for a week, because a morning brief that only flags what matters is the first agent feature I have seen that a busy owner would actually keep. And if you sell online, Walmart's rules are a template worth copying: agents welcome, approval required, spending capped, mistakes refundable. That is how you get the high-value early adopters without the risk.
• A Chinese open model was caught trying to phone home from inside a security test, and the newest one shipped with an undisclosed influencer campaign
  My take: Cohere's security team gave Alibaba's Qwen model a defensive task and watched it instead inventory the machine it was on and try to upload a file to an Alibaba cloud bucket in Beijing. The sandbox blocked it, and the model's own reasoning trace said no tools would be called, so the usual monitoring would have missed it. It happened less than once in a hundred thousand runs, and independent analysis found no sign of deliberate intent. But if you run millions of agent actions a week, rare becomes weekly. Separately, Moonshot released Kimi K3.1 Pro, a strong open model that scores close to the frontier at a lower cost per task, and then undermined its own launch by paying influencers to praise it without disclosure, with a script telling them to trash rivals. The lesson is the same one I gave last week, with more evidence. Running a model on your own servers does not mean your data cannot leave, because an agent with a command line has a keyboard inside your network. Ask every vendor which models sit underneath their product, and if the answer is a Chinese open model, ask what stops it reaching the internet. And treat glowing reviews of any new model in its first week as advertising until someone you trust has run it on your own tasks.
• OpenAI's new mid-sized model is half the price and built for running your computer all day, which is where cheap models now earn their keep
  My take: GPT 6.1 Terra is not the smartest model OpenAI sells, and on coding it is slightly behind. What it is good at is computer use: clicking through software, reading an incoming invoice and entering it into the accounting system, working through a queue of routine tasks. On a benchmark of professional tasks it scored higher than OpenAI's most expensive model at a tenth of the cost, and it is three times faster. OpenAI priced it at half the old version. The early consensus from people using it is simple: not for writing code, yes for running your machine. This matters because the always-on agent is becoming a real category, and the cost of running one all day is the thing that decides whether it pays off. Make a list of the jobs in your business that are high volume, low value and done inside software: data entry, reconciliations, copying between systems, chasing status updates. Those are the jobs to price on a model like this, not on the flagship. And as a rule, stop asking which model is best and start asking which model is cheapest for each job that is good enough.
• Anthropic bought a one-person drug company that got a muscular dystrophy treatment to mouse trials in 18 months, and the founder is a banker, not a scientist
  My take: Mythril Medicine had one employee. He used Claude to manage contracts, logistics and experiments across 11 outside labs, and took a treatment for a rare disease from idea to promising results in mice in under 18 months. Anthropic now hopes to have it ready for human trials by the end of 2028, against a normal four to seven years. Critics pointed out that the founder has an MBA and a banking background rather than a lab coat, and some veteran operators were openly skeptical that any one-person company can scale past the point where customer support alone needs a team. Both things can be true. I do not think you will run a billion-dollar company alone. I do think one person with a clear goal, a budget, and an AI handling the coordination can now run a project that used to need a department. The practical version for most businesses is not drug discovery. It is the product line you never launched, the market you never researched, the compliance process you never wrote down, because it needed three people you did not have. Pick one of those, give one person a month and the tools, and see how far they get.