Anthropic adds restrictions to Claude Code to prevent game recreations

New server-side restrictions

Anthropic recently introduced a new set of instructions for its Claude Code tool, specifically targeting the Opus 5.5 model. These instructions are designed to prevent the AI from being used to recreate specific existing video games. The goal is to protect intellectual property from being copied by users who might try to disguise the recreation by changing the name of the project.

The update was implemented as a server-side configuration flag known as 'tengu_heron_brook'. Because this change is delivered remotely by Anthropic, it does not require a traditional software update from the user. Currently, there is no local setting available for developers to disable these specific restrictions if they find them inconvenient.

Addy Osmani, a member of the technical staff at Anthropic, confirmed the addition of these two sentences to the model's prompt. He explained that using the AI to infringe on intellectual property violates the company's usage policy. The company intends for the tool to still support original games, projects within established genres, and work based on a user's own designs.

Impact on independent developers

While Anthropic aims to stop piracy, some independent developers are reporting that the tool is being over-zealous. One developer from Conflict Simulations LLC reported that Claude Code began refusing commissioned work that the business had successfully completed for months. These projects involved creating personalized digital modules of board wargames for individual clients.

The developer noted that these modules were built using code written from scratch and materials provided by the clients. Although the projects were for individual use rather than public distribution, the AI began refusing the requests once the developer asked if the new restrictions applied. The developer clarified that while they owned the physical games, they did not have formal authorization from the publishers to adapt the copyrighted artwork or text.

This situation highlights a gray area in AI safety where the model may struggle to distinguish between authorized adaptation and unauthorized copying. The developer eventually had to escalate the issue to human support staff after the automated system initially denied the request without exception. This friction suggests that the current automated filters may lack the nuance required to evaluate complex, commissioned creative work.

The context of AI copying

These new restrictions come at a time of growing concern regarding how AI can be used to imitate existing games. For instance, the independent publisher Alliance Arts recently canceled the public release of a demo for a game called Drunken Goddess Reflux. The studio expressed concerns that the demo could be used to create AI-assisted imitations of their unfinished title.

The studio did not report that an AI-generated copy of the game actually existed, but they chose to limit public exposure to protect their work. This move reflects a broader trend where smaller developers are seeking tighter controls over how their games are shown and distributed. They are trying to prevent their unique mechanics and aesthetics from being ingested into training data or used as blueprints for clones.

Anthropic has not yet released specific data on how often these new instructions trigger a refusal. There is currently no public information regarding how many of these refusals involve projects that were not attempting to reproduce existing games. The company has not provided a detailed set of criteria for how the AI makes these distinctions in real-time.

Challenges in enforcement

A major challenge for Anthropic is determining how to verify whether a developer's project is truly original or a disguised copy. Because the instructions are designed to block recreations even if the project is renamed, the AI must rely on its internal understanding of game mechanics and assets. This can lead to 'false positives' where legitimate creative work is flagged as a violation of policy.

Anthropic has invited developers to report cases where they believe legitimate projects are being incorrectly refused by the system. However, the company has not yet established a formal verification process to handle these appeals for commissioned work. This leaves a gap in how the company will balance the protection of existing titles with the needs of creators who are building new content.

As the AI landscape evolves, the tension between copyright protection and creative freedom remains a central theme. Developers will likely need to be more specific in their prompts to ensure the AI understands the original nature of their work. For now, the 'tengu_heron_brook' flag remains a primary tool for Anthropic to manage these risks at the server level.