Developers have reported a wave of refusals from Claude Code since late October, with some saying the AI coding assistant will not help with projects they believe they are entitled to build. The complaints center on what developers describe as a server-side instruction that blocks work related to games. The restriction appears to be applied broadly, without an obvious way to distinguish commercial piracy from commissioned work for a client who owns the underlying game.

Anthropic has not announced a policy change, and the Claude Code changelog does not mention one. There has been no official statement explaining whether a new game-related rule exists, how it is enforced, or whether professional developers can seek an exemption. That silence has left developers to compare notes in public issue reports and support threads, trying to understand why some sessions refuse tasks that previously worked.

Key facts at a glance

• Developers began reporting Claude Code refusals on game-related work in late October.
• A studio owner who builds personalized digital versions of board wargames for clients who own the games said Claude Code began citing a server-side instruction against reproducing games around 27 October.
• The developer filed a public issue on 30 October and later appealed to Anthropic, asking it to target actual copyright infringement and provide a verification path for professionals.
• Anthropic has not confirmed a new game policy, and its changelog does not mention one.
• Separately, developers reported cyber-safety blocks on anti-cheat code and a security review of a game.

The most detailed account comes from a developer who runs a small studio that creates personalized digital versions of board wargames. The clients own the physical games, and the developer describes the work as commissioned professional adaptation rather than piracy. Around 27 October, the developer wrote in a public issue, Claude Code began getting what they called “a server-side instruction” to refuse game work. Sessions then refused the work and cited an instruction against reproducing games.

The developer objected to the breadth of the rule. “A blanket category rule against ‘games’ cannot tell the difference between piracy and commissioned professional work for a game’s owner,” the developer wrote. That complaint captures the central tension: AI safety rules often operate at the level of categories, while real development work is defined by contracts, licenses, ownership, and context. A rule designed to prevent unauthorized cloning can also catch legitimate adaptation, accessibility work, preservation, and client-specific projects.

How the refusal appeared

The developer’s account includes a puzzling detail. In an update, they said sessions kept doing the work after the rule reportedly arrived. The refusals began only after they asked Claude to check whether the rule affected their projects. The refusals then continued in new sessions on the same code. One commenter suggested that the project’s memory or Git history might now carry the refusal, and the developer said they would try to clear it.

That sequence points to a broader problem with AI coding assistants: policy enforcement can be inconsistent, context-dependent, and difficult to audit. A model may comply with a task until a user asks about a restriction, at which point the assistant becomes more cautious. Once a refusal is embedded in conversation history, project memory, or repository context, it may persist even when the underlying task has not changed. Developers can be left troubleshooting not only their code but also the assistant’s interpretation of policy.

The developer also said Anthropic’s support bot told them the policy covers all customers, with no exception. Both support tickets were later passed to staff, and the developer filed an appeal. They asked Anthropic to aim the rule at real copyright infringement and to offer a verification path for professionals. That request echoes how security researchers often need a way to prove their work is legitimate before gaining access to restricted capabilities.

A divided developer response

Not everyone on the public thread agreed with the complaint. One commenter said that unless the work is for the rights holders, it is no different from any other reverse-engineered port. That view reflects a real debate in game development and preservation. Some projects are clearly licensed; others operate in gray areas; still others are outright infringing. The difficulty is that a coding assistant cannot easily verify the difference from a chat prompt alone.

For professional developers, the stakes are practical. A studio may be paid to adapt a game for a client, create a digital companion, or rebuild mechanics for accessibility. Those projects can involve proprietary rules, art, or code. If an AI assistant refuses the entire category of games, the developer must either abandon the tool, restructure the workflow, or seek a manual exemption. Each option adds cost and uncertainty.

The dispute also raises questions about how AI providers should document policy changes. Developers rely on changelogs, support documentation, and consistent behavior to plan their work. When a restriction appears without a public announcement, users cannot tell whether it is a bug, a temporary filter, a model behavior, or a deliberate policy. That ambiguity can damage trust even if the underlying safety goal is reasonable.

Separate cyber-safety blocks

Other developers reported a different kind of block this month. Claude Code’s cyber safeguards flagged anti-cheat code for one developer’s game and a security review of another developer’s own game. The timing of the two kinds of restriction may be a coincidence, but the overlap has added to confusion. Game development often involves anti-cheat systems, security reviews, and code that resembles the kind of work cyber-safety filters are designed to catch.

Anthropic says its generally available models have conservative cyber safeguards that block most cyber work. On 6 October, the company expanded its Cyber Verification Program, which lifts some of those blocks for vetted security teams. Members of Project Glasswing, an earlier programme for its Mythos model, move into the top tier. The existence of that program shows that Anthropic already has a model for distinguishing vetted professionals from general users in at least one sensitive domain.

That model could be relevant to the game dispute. If Anthropic can verify security teams, it could potentially verify game developers, publishers, or rights holders. A verification path would not resolve every copyright question, but it could reduce false positives for legitimate work. The developer’s appeal asks for exactly that kind of mechanism. So far, Anthropic has not said whether such a path will be offered for game-related work.

Copyright pressure and platform risk

The refusals come amid broader copyright pressure on AI companies. Anthropic faces copyright claims, including a suit over 274 Elvis Presley songs filed in October. Those disputes concern training data, reproduction, and licensing, and they help explain why AI providers are cautious about requests that appear to recreate copyrighted works. A coding assistant that helps build a game may be asked to reproduce mechanics, assets, or code that belong to someone else.

At the same time, copyright law and licensing practice are more nuanced than a blanket category ban. Game developers routinely work with owned assets, licensed engines, and client-approved materials. A studio may have a contract that authorizes adaptation. A client may own the game and commission a digital version for personal use. A preservation project may rely on precise reproduction for archiving. These cases require context that a short prompt may not provide.

Separately, a tampered Claude Code extension reached users of another AI coding tool on 29 October. That incident highlights a different risk: developers are installing extensions and integrations that can affect their code and credentials. It is not directly connected to the game refusals, but it adds to the broader sense that AI coding workflows are becoming more complex and harder to secure.

What developers want next

The developer behind the appeal wants Anthropic to narrow the rule to real copyright infringement. That would mean looking for indicators of unauthorized reproduction rather than treating every game-related request as suspect. It would also mean giving professionals a way to document their rights or client relationships. Without that, developers say, the rule functions as a blunt instrument.

Anthropic has not said whether the game rule exists or will get an exemption path. The changelog does not mention one, and there has been no official statement. For now, developers are left with public issue reports, support responses, and their own experiments. Some may move to other tools; others may keep testing whether new sessions, cleared memory, or adjusted prompts avoid the refusal.

The episode illustrates a larger challenge for AI coding assistants. Safety filters are necessary, especially when tools can generate code that touches copyright, security, or proprietary systems. But filters that are opaque, broad, and inconsistently enforced can block legitimate work and push developers away. The dispute over game refusals is not only about one category of software. It is about how AI providers communicate policy, verify professional use, and build appeals that users can trust.

Source: TNW | Anthropic News