Software got cloned in a weekend and game studios pivoted, a coding tool leaked every customer's API keys, and businesses now have real alternatives to renting intelligence

A developer rebuilt Photoshop with Claude and Epic banned handwritten code, which shows the new models can take apart and remake software fast enough to change business models. Meanwhile stolen API keys ran up bills for thousands of users, Microsoft and Meta released models that let companies own the result, and the fight over AI budgets and AI personality moved from theory to practice.

• A developer cloned Photoshop with Claude, Epic banned handwritten code, and Xbox opened a lab instead of calling the lawyers
  My take: The newest models can take a compiled program apart, document what it does, and rebuild it from scratch, which used to take a legal team and two groups of engineers. One developer did it to Adobe's suite with a Claude subscription and Adobe's stock dropped 3 percent on the news. Gamers are doing the same to console titles. Epic responded by banning handwritten code and moving an entire Disney game to AI-first production, Xbox created a lab to experiment in the open, and a private equity firm bought Bumble on the thesis that a small AI-first team can run it better. Software subscriptions are not about to vanish, because most businesses pay for the support and the accountability, not the code. But the moat around any product that is just software got a lot thinner this month. If you sell software, ask what a customer still gets from you that a clone cannot give them: your data, your service, your integrations, your brand. If you buy software, your negotiating position just improved, and a mid-size vendor that will not move on price is now easier to replace.
• A coding tool leaked its users' API keys to an agent swarm, Meta's Muse exposed business loan applications, and Bank of America hired 1,000 contractors to patch bugs
  My take: OpenCode, a popular open-source coding tool, had its database of stored customer API keys stolen. Attackers used those keys to run up bills in the hundreds and thousands of dollars per user, generating tokens to sell or to train their own models. The founder said the attackers were agents running around the clock, and he needed agents of his own to keep up. In the same week, security researchers found that Muse's conversation sharing links were easy to guess and pulled up 6,800 of them, including a small business loan application and a company's internal revenue review. And Bank of America disclosed it has 1,000 contractors fixing bugs that new AI models can now find. Three actions. First, treat every service that holds your API keys as you would treat a bank, and give each tool its own key with a spending cap so a leak costs you a fixed amount. Second, never paste customer or financial information into an AI chat you might share, and tell your staff the same. Third, if you run any software that faces the internet, assume someone with an AI model will probe it this year, and ask your developer or IT provider what they have done about it since the summer.
• Microsoft and Meta released models you can own, and the US government banned DeepSeek from every federal system
  My take: Microsoft announced a new base model it says beats OpenAI's current model at a quarter of the training cost, and it sells a service that tunes it on your data so the result is yours. Meta released Muse Spark Open, the most capable American open-weight model so far, with full downloads and first-party support to run it on Apple hardware. The same week the Department of Homeland Security banned DeepSeek from all federal systems, including versions hosted on American servers and products built on top of it. Put those together and the picture is clear. If you want AI you control, you no longer have to choose between a Chinese model and nothing, and if you sell to government or to anyone in a defence supply chain, you need to know which models your vendors built on. For most small businesses the practical step is smaller. Ask each AI vendor you pay which models they use underneath and whether they can swap. A vendor that cannot answer has locked you in without telling you.
• AI budgets are moving from unlimited to earned: Devin now raises your token limit when you prove you use it well
  My take: Companies swung from unlimited AI spending early this year to hard monthly caps once the bills arrived. Cognition, which makes the Devin coding agent, has built a middle path. It estimates how many human hours each session saved, scores how efficiently an engineer used their budget, and raises the limit automatically for good stewards, like a credit card limit. It also added a coach that flags common waste, such as vague requests or an expensive model for a trivial task. Developers hated it, and some of the criticism is fair, because hours saved rewards building complicated things. But some version of this is coming to every business, because AI spend is now a real line item. You do not need their tool to copy the idea. Give your team a clear monthly budget per person, make usage visible, and when someone shows they get real work out of it, give them more rather than less. The people who are good at this will be your most productive staff within a year, and starving them to save a few hundred dollars is a bad trade.
• Meta's Muse hit 10 million US downloads in 40 days, and the labs are now betting personality matters as much as intelligence
  My take: Muse grew faster than any app except a handful of mobile games, and the people who use it keep saying the same thing: it is fun, and the mascot made it something people screenshot and share. OpenAI's rival agent is described as capable but dry. A new independent benchmark now scores models on 30 personality traits, from warmth to directness to filler. The business point is that as every model gets good enough, how it feels to work with decides which one people actually use, and that applies inside your company too. If your team has quietly stopped using the tool you bought, the problem may be the tool's manner, not its brains. Let people try two or three options on the same tasks and keep the one they come back to. And if you are putting an AI in front of customers, test its tone with real customers before you launch, because a correct answer delivered like a brick wall still loses the sale.
• A stock research firm designed scorpion antivenom candidates with Claude for under 200,000 dollars, and the lesson is about what your business can now attempt
  My take: A financial research firm wanted to know if AI drug discovery was real before investing in the sector. Instead of interviewing experts, they used Claude to design proteins, paid a lab to make and test them, and found 17 that bind to scorpion toxins. The design took 64 hours of compute. The lab work took 39 days, and any real treatment is years away, so the physical world is still the slow part. They are not becoming a biotech company. They did get a far better answer to their investment question than any survey could give them. That is the pattern to notice. AI does not just make your current work faster, it makes it affordable to do things adjacent to your business that you would never have attempted. The accountant can build the client dashboard, the retailer can run the supplier analysis, the agency can prototype the product instead of pitching it. Pick one question about your business you have always answered by guessing or asking around, and spend a week seeing whether you can now go and find out for yourself.