TL;DR: Anthropic quietly added a new instruction telling Claude to stop recreating existing video games, even when someone asks it to “reimagine” the game or gives it a new name. The change wasn’t announced. It surfaced through an X post, and Anthropic staffer Addy Osmani confirmed it, tying the move to a wave of AI-decompiled AAA games, including Halo: Combat Evolved, Call of Duty: Modern Warfare 2 and Grand Theft Auto V, that have been showing up playable in browsers.
The discovery started with an X user who goes by chasmmmmmmmmmmm, who spotted new language buried in Claude’s instructions. The relevant lines, as quoted in the post: “Do build original games, games in a genre, and games from the user’s own designs and assets.” And then the restriction: “Do not build a recreation of a specific existing game, even if it is renamed or called original.”
Addy Osmani, who works on Anthropic’s technical staff, responded on X to confirm the change was real and intentional. He said the company wants to steer Claude “away from recreating someone else’s specific game because using Claude to infringe IP is against our usage policy.” He added that original games, genre games, and games built from a user’s own designs and assets should still work fine, and that Anthropic will “continue to iterate to ensure we’re striking the right balance.”
That’s a narrower fix than it might sound. Claude can still help someone build a platformer, a roguelike, or a card game from scratch. What it’s now supposed to refuse is a straight clone of a specific, identifiable title, dressed up with a different name or a coat of “original” paint.
The timing tracks with something that’s been spreading across game-dev and AI communities for weeks: AI tools decompiling classic and modern AAA titles and rebuilding them as browser-playable projects. The article that broke this story points to specific examples making the rounds, including a vibe-coded Halo: Combat Evolved recreation, a Call of Duty: Modern Warfare 2 project circulating on Reddit, and a Grand Theft Auto V rebuild covered by IGN. None of those projects are official, and none of the studios behind the originals, Bungie, Activision, or Rockstar, appear to have filed anything over this specific wave. But the pattern is obvious enough that an AI company whose tools can produce this kind of output doesn’t need a lawsuit to see the exposure coming.
Anthropic has plenty of reason to move early here. Last year the company agreed to a record settlement of $1.5 billion with a class of roughly 500,000 authors over training data, one of the largest copyright payouts in AI history. The Elvis Presley estate has also sued Anthropic over similar claims. Those cases were about what Claude was trained on. A wave of recognizable game clones would be a different, arguably messier problem: it’s about what Claude helps users build, with a named, trademarked, copyrighted product sitting at the other end of the chat.
Anthropic has been tightening Claude’s guardrails on multiple fronts this year, not just copyright. The company’s own alignment research, including its Mythos 5 report on why one of its models attacked a real GitHub project, shows a pattern of Anthropic publicly documenting and patching behavior it doesn’t want Claude exhibiting, rather than waiting for outside pressure to force the issue. The video game instruction fits that same playbook: quiet, specific, and reactive to a real pattern of misuse rather than a hypothetical one.
It’s also a reminder that the boundaries around what AI coding tools are allowed to produce are still being written in real time, often in response to what users actually do with them rather than what companies anticipate. Claude isn’t the only AI coding assistant dealing with scrutiny lately. Just weeks ago, a fake Claude Code extension on the Cursor marketplace planted SSH backdoors on up to 5,000 developer machines, a separate incident but part of the same broader story: AI coding tools have gotten good enough, and popular enough, that both the companies behind them and the people trying to abuse them are moving fast.
There’s no public date attached to when this instruction went live, and Anthropic hasn’t published a blog post or changelog entry about it as of this writing. The confirmation exists only as a reply on X from one staffer, not a formal company statement. That’s a thinner paper trail than you’d expect for a policy change this specific, and it leaves some real questions open: Does this apply retroactively to conversations already in progress? Is it enforced by a system prompt, a classifier, or both? And how does Claude tell the difference between “a genre game” and “a recreation with the serial numbers filed off,” which is exactly the gray area the instruction is trying to close?
For now, the practical effect is simple. Ask Claude to help build an original platformer, and it should still help. Ask it to rebuild Super Mario Bros. with a new coat of paint, and it’s supposed to say no. Whether that holds up against a determined user with a clever prompt is a separate question, and probably the next chapter of this story.
It’s worth sitting with how fast this trend moved from novelty to policy problem. A year ago, “vibe coding” a playable game clone from scratch in an afternoon was a party trick people posted for likes. Now it’s common enough that an AI lab felt the need to write a specific carve-out into its model’s behavior, rather than relying on a general copyright clause buried in its terms of service. That’s the broader pattern across every major AI coding tool right now: usage policies read like static documents, but the actual behavior underneath them keeps getting patched week to week as real-world prompts expose gaps nobody wrote a rule for in advance. Game studios watching this unfold don’t need to sue anyone yet. They just need Anthropic, OpenAI, and everyone else building coding assistants to keep noticing these gaps before a court does it for them.
The bottom line
This is a small, specific fix to a real problem, and Anthropic deserves some credit for making it before a studio’s lawyers forced the issue. But announcing a policy change through a staffer’s X reply, rather than a changelog entry users can actually find, is a pattern worth watching. Anthropic is already fighting a $1.5 billion copyright settlement and a lawsuit from the Presley estate. Quietly patching Claude’s behavior without quietly patching how it communicates those changes is the kind of gap that tends to come back around.
The Teck Hustlers Editorial Team covers technology, business, finance and gaming news for readers who want the story behind the headline. Every article is researched, written and edited by our in-house team, checked against company announcements, filings and reputable reporting, and updated when new facts emerge. We focus on what each development means for founders, professionals, investors and everyday users. Have a tip or a correction? Reach us through our contact page.
