Skip to content

AI News · Edition 83 ·

Your AI tool's rules can change overnight. Plan for it

Anthropic reins in Claude over game copies, and some legitimate work gets caught. Plus: hijacked ChatGPT accounts and AI-assisted crooks.

8 stories · sourced from Tom's Hardware, Hacker News, CyberSecurityNews, Engadget

Abstract illustration accompanying Your AI tool's rules can change overnight. Plan for it
Abstract illustration, generated with AI. It represents the idea, not the event.

The one that matters

Anthropic reins in Claude on copying games, and reports say legitimate developer work is caught too

Tom's Hardware ↗

What happened

Anthropic, the company behind the Claude AI assistant, has changed Claude so that it pushes back when people ask it to clone or remake existing video games. At least seven outlets covered the change. PC Gamer says it follows weeks of Steam, the largest PC game store, filling up with low-quality AI-made copies. IGN reports the update was made to Claude Opus, Anthropic's top model. TweakTown says it also covers Claude Code, the version programmers use to write software.

Tom's Hardware goes further. It reports that the restrictions also cover reverse-engineering, which means working out how an existing program works from the outside. It also reports that legitimate developers and basic troubleshooting tasks are getting caught. The outlets describe the change differently. PC Gamer's headline describes an effort to “steer” Claude away from cloning, while Tom's Hardware's says Claude “blocks” it. Nobody yet knows how firm the line is or how often it catches the wrong work.

What it means for your business

Most readers don't make games, so the games are not the point. The point is that an AI tool you or your developer rents can change what it will do overnight, with no new contract and no notice to you. If your software is built or maintained with AI help, which is now common for agencies in Pune and Toronto alike, that tool's rules are part of your supply chain.

The work most likely to run into a reverse-engineering rule is quite ordinary for small businesses. Take a 600-student school whose fee software came from a vendor that has since disappeared, and which now wants its records moved to a new system. Or take a workshop with 40 staff running a 15-year-old inventory program whose source code nobody has. Working out how that old program stores its data counts, technically, as reverse-engineering. The reports do not say jobs like these are being refused. But if basic troubleshooting is getting caught, as Tom's Hardware reports, you should expect some friction.

The usual cost is time, not a dead end. When a developer's AI tool refuses a step, they switch tools or do that step by hand. A data migration quoted at three days might take four or five. On a fixed-price job, that kind of delay often turns into an argument about who pays.

What it does not mean

  • It does not mean Claude is now useless for business software. The change is aimed at copying games. Nothing in the reports suggests it affects building a booking page for a 30-bed clinic or a daily sales report for a shop. The “sweeping” label comes from developers who ran into refusals. These are early reports, and nobody knows yet how widespread the problem is.
  • It does not fix the flood of AI-made game copies. Anyone calling it a cure is overstating it. This is one company's model, other tools exist, and steering a model away from something is not the same as blocking it.
  • It is not a change in the law. It is a product decision. What you are legally allowed to do with your own old software, or with a competitor's, is the same as it was last month.

What a sensible owner should do this month

  • If you pay a developer or agency, ask which AI tools they depend on, and what happens to your deadline if one of them refuses a task or changes its rules. A good answer names a backup plan. A shrug is a warning sign.
  • If you have an old system you plan to move off, say so clearly when you ask for quotes. Ask the developer to test the hardest step, getting the data out, before the price is fixed. One day of testing can prevent a week of dispute.
  • If staff use AI tools in daily work, make sure no routine task depends on a single tool with no manual backup. A tool's rules can change, but your fee deadline and month-end close won't move.
  • If none of this applies to you, you have nothing to do. Just remember that AI tools are rented, not owned, and the company that rents them to you sets the rules.

Also worth knowing

  1. A widely discussed write-up describes taking over ChatGPT accounts

    A post titled “We Took over ChatGPT Accounts” was among the most discussed items among developers. Early coverage doesn't make clear how it was done or whether the flaw has been fixed. Either way, anyone who gets into a ChatGPT account where staff have pasted fee records, customer lists or patient notes can read all of it, so turn on two-step login and set a plain rule about what never goes into a chatbot.

    Hacker News ↗

  2. Malware is using AI to sort its victims by how much they're worth

    A report describes malicious software that uses AI to rank the people and businesses it infects by how much they're worth. Owners of small clinics and shops have often assumed they're too small to be worth a criminal's time, but that matters less when sorting thousands of victims costs criminals almost nothing. You don't need to buy anything new: offline backups, two-step logins and prompt software updates are still your defence.

    Hacker News ↗

  3. An Accenture contractor allegedly connected a Claude AI agent to a secure Pentagon network

    This is an allegation reported by two outlets, and the full facts are not settled. The lesson for a 40-person office is simple: an AI agent can act on any system it is connected to, and one well-meaning employee can connect it to systems it should never touch. Decide who is allowed to connect AI tools to email, accounts and records, and write that rule down.

    CyberSecurityNews ↗

  4. Anthropic's Claude Duet agent is arriving in stages. Go slowly when it reaches you

    Anthropic is rolling out Claude Duet, an agent that one report says links six work apps, to users in stages rather than all at once. If it shows up on your team's accounts, try it on low-stakes tasks first, and remember the Pentagon story above: an agent that can read your inbox and files can read everything in them. Keep it away from patient, student and payroll data until you understand what it does with that data.

    Engadget ↗

  5. Cisco gives one engineer unlimited AI usage to rewrite old code

    Cisco has reportedly given an engineer an unlimited AI budget to rewrite old software, and developers paid close attention. It suggests that large companies see AI as a practical way to update old systems, the same problem a factory running a 20-year-old ERP has. But “unlimited” at Cisco's size tells you nothing about what it would cost you, so if you're tempted, pay for a small fixed-price trial on one part of your system first.

    Hacker News ↗

  6. US tech stocks fell sharply on fears about what AI costs

    US tech shares fell sharply as investors worried about the cost of AI. For a shop, school or clinic, nothing changes this month. AI companies under cost pressure may raise prices or cut free plans, so choose monthly plans over long prepaid AI contracts for now.

    Hacker News ↗

  7. A 2025 study on AI and critical thinking is being discussed again. It relies on self-reports

    A 2025 study on how generative AI affects critical thinking is being discussed again. Its own title points to its main limit: it records what people say about themselves, not their tested ability. Schools and training managers should treat any headline about it as a reason to ask questions, not as proof either way.

    Hacker News ↗

How this briefing is put together

Every morning we read the day’s AI announcements and reporting from the companies themselves and from the technology press, then pick the handful that actually change something for a working business. The analysis is ours and it is written for owners and managers, not engineers. Every story links to its original source — read them, and disagree with us where we’ve got it wrong.