Anthropic Quietly Blocks Claude From Reverse-Engineering Games

What Anthropic actually changed

There was no blog post, no policy update banner and no developer changelog. The restrictions surfaced through users noticing that Claude had started declining certain requests.

Claude has begun refusing or heavily limiting requests to decompile, mod or reverse-engineer game code, according to reporting from Tom's Hardware. Anthropic has not said publicly why, and the company has not published a revised usage policy explaining where the new line sits.

That silence is the awkward part. A model that will happily explain how a game engine works in the abstract now balks when the same question is framed around an actual shipped game. Users have no reliable way to know which side of the line a prompt falls on until it gets refused.

Why Anthropic is blocking Claude from game code

The timing points at copyright pressure rather than anything specific to gaming.

The last year has been rough for game publishers. GTA 6 leaks went viral and pulled Take-Two Interactive into subpoenas aimed at Microsoft and Discord. Then the modding scene discovered that AI models make short work of the tedious parts of decompilation, and the results stopped looking like hobby projects.

Modders dropped Mario into Elden Ring and Spider-Man into Batman: Arkham City. Ports of World of Warcraft, Call of Duty and GTA: Vice City turned up playable in a browser, with no involvement from the rights holders. Then came the PlayStation 5 jailbreak, which produced backward compatibility with older PlayStation games, cross-platform play and a way to run Halo 2 and Zelda: Breath of the Wild on Sony's console.

Any one of those could turn into a lawsuit with damages large enough to make a general counsel sweat. Blocking the tool at the source is cheaper than litigating later.

The filters are catching the wrong people

The restrictions are broad enough that legitimate work gets caught in them.

Game developers use Claude and similar models for ordinary tasks: reading a crash log, untangling a build error, understanding a third-party library. The new limits reportedly interfere with some of that, and with basic troubleshooting for modders working on projects nobody would call piracy.

There is a real irony here, and Tom's Hardware's Bruno Ferreira put it plainly: frontier AI companies that trained on the world's information and art are now the ones drawing a copyright line. That does not make the concern fake. It does make the position harder to defend in public, which may explain why Anthropic said nothing.

The catch: someone else's model will do it

A restriction only works if every competitor follows it.

Anyone determined to decompile a game can move to a model that has not adopted the same filters, or run an open-weight model locally where no provider can refuse. The least restrictive assistant becomes the most useful one for exactly the work Anthropic is trying to discourage.

That is the uncomfortable incentive at the center of this. Anthropic can make a principled call, and the people it most wants to stop are the least likely to be affected by it. It is the same pattern we have seen with AI agents that can complete complex tasks—like the one that beat Factorio in 13.5 hours—where capability and control are always racing each other.

What it means for you

The change applies to Claude wherever you use it, so US, UK and EU subscribers on claude.ai and the API are all affected the same way. There is no regional version of the policy and no opt-in for verified developers.

If your work involves game code, the practical advice is to keep a local open-weight model around for the tasks Claude now refuses. Anthropic has not announced any developer exemption program, and until it does, assume the filters stay.

Anthropic is right that AI-assisted decompilation is heading somewhere the games industry will not tolerate. But a quiet filter on a paid tool is the worst of both worlds: it annoys the developers who play by the rules and barely slows the ones who don't.