Anthropic has added two sentences to the instructions for using Claude Opus 5.5 in its AI coding tool Claude Code, asking it not to recreate other people's existing games. Addy Osmani of Anthropic revealed the change in an X post on October 30, 2026 (October 31 in Japan). He said the change applies to code generation the company's existing policy that prohibits using Claude to infringe intellectual property rights.
The change is not a ban on game development in general. Anthropic says developing original games, games that belong to a particular genre, and games based on a user's own designs or assets should remain possible. Meanwhile, reports on GitHub say that Claude Code has refused work on a game-related contract project, suggesting the restriction may be reaching situations it was not meant to cover. To understand the change, it helps to distinguish recreating an existing game from making an original one, and to separate contract work and its rights questions from the different framework that applies to reverse engineering in security work.
What do the two added sentences restrict?
Osmani explained that Anthropic added two sentences to the Claude Code prompt for Opus 5.5 so that it avoids generating code that recreates another party's specific game. He said that using Claude to infringe intellectual property rights has always violated the company's usage policy, and that this change reflects that policy more concretely in responses related to game development.
Osmani also stated clearly that users should still be able to develop original games, games that incorporate features of an existing genre, and games that use assets the user designed and created. For example, building an original game while drawing on mechanics common to a genre is treated differently from recreating another party's existing work itself. The new instruction is not meant to prohibit requests to "make a game" in general; its purpose is to avoid recreating specific works.
In its report, the game development outlet Game Developer said that X user Chasm discovered the new instruction in Claude Code. The instruction reportedly says not to recreate a specific game even if its name is changed or it is described as an original work. Chasm said the instruction appears to have been added around October 27, and that no setting for users to disable it could be found. Asked by Game Developer, an Anthropic spokesperson pointed to Osmani's post, whose explanation confirmed the change.
However, what Anthropic described was a change to the Opus 5.5 prompt in Claude Code. It has not said the same instruction applies to all Claude models or products. Nor did it say the model itself was retrained, or announce new terms of service. The change is positioned as an alteration to the instructions given to the model during code generation.
The table below summarizes this change alongside the separate security-access program Anthropic offers.
| Use | Anthropic's explanation | Scope and conditions | Caveat |
|---|---|---|---|
| Recreating another party's specific game | Claude is guided to avoid recreation | Instruction for Opus 5.5 in Claude Code | Application to all Claude models has not been confirmed |
| Original games, genre-based games, and games made from one's own designs and assets | Expected to remain possible | Development that Osmani explicitly named as continuing | Not a guarantee that every individual request will be accepted |
| Defensive malware analysis | Reverse engineering is included as a target use | Defense Access in the Cyber Verification Program, with conditions such as qualifications and authorization | Not an exemption system for game development |
The difference can be seen from Osmani's October 30 post and from Anthropic's expansion of the Cyber Verification Program, announced on October 6. Even for similar coding or analysis work, the conditions that apply differ depending on the software involved, the purpose of use, and access permissions. The fact that a particular request was refused also does not mean the work was judged illegal.
A contract-work refusal report: the claim of being "licensed" was corrected
In Issue #106167 on Claude Code's official GitHub, a user who does game-related contract development reports being refused work after the new instruction. The poster, lerugray, says they run a business called Conflict Simulations LLC, producing paid digital modules of board wargames owned by customers, for those customers' personal use. They say they write original code based on materials supplied by the customer, and do not sell or distribute the finished modules to the public.
There is, however, an important correction to this report. On October 30, the poster admitted that the term "licensed," which they had originally used, was not accurate, and stated clearly that they are not producing the modules under a license from the publishers. The explanation that the customer owns the game and supplies the materials is different from an explanation that the copyright holder or publisher has granted permission to copy, adapt, or otherwise use the work. This case therefore cannot be treated as one in which publisher-authorized development was refused.
At the same time, the absence of a publisher license does not by itself establish that each act of production is illegal. The legal assessment varies with the acts involved, the materials used, the rights relationships, and other factors. What this report establishes is limited to the poster's account that they confirm the customer's ownership of the game before doing contract work, and their claim that Claude Code began refusing that work.
The circumstances in which the refusals began also deserve attention. According to the poster's follow-up, they were able to keep working as before for some time after October 27, when the new instruction is said to have been introduced. But after they asked Claude itself about the published instruction, it began refusing the same development work. They also report the same problem on multiple machines and on both paid Max and Team plans.
If this account is correct, work did not stop simply because the new instruction was introduced; having the model check its own criteria may have been connected to the refusals. However, there has been no independent reproduction and no explanation of the internal cause from Anthropic, so it cannot be concluded that the two added sentences directly caused any individual refusal. The poster asks for clarification of the restrictions, a mechanism to appeal false positives, and a system for verifying authority to do contract work.
In its help article on outputs blocked by content filtering, Anthropic has previously explained that a refusal does not necessarily mean the company has judged the content inappropriate. Because the mechanism is designed to prevent copying and reproducing existing copyrighted works, refusals can occur even when the user has no intent to copy. Still, this general explanation alone cannot establish that the refusal in the contract-work case was a false positive. A report that a refusal occurred and whether the reason for it was legitimate need to be considered separately.
Has reverse engineering been banned across the board?
Separately from the instruction to avoid recreating games, there are concerns that Claude has stopped handling reverse engineering in general. But on October 6, 2026, Anthropic expanded the Cyber Verification Program (CVP) and explicitly included defensive reverse engineering among its target uses. Anthropic has not uniformly banned the work of analyzing how programs operate for all Claude users.
The CVP is a program that grants security professionals who meet certain qualifications and conditions access to advanced cybersecurity capabilities and to access with fewer restrictions than usual. The expanded program has several access categories. Among them, Defense Access covers malware reverse engineering, incident response, and vulnerability analysis and verification. Eligible applicants include not only security teams at companies, universities, and government agencies but also individual researchers with a certain track record. Use, however, is subject to conditions such as verification of qualifications and required security controls, and not everyone can use the same capabilities without restriction.
Anthropic also explains that, even with its generally available models, tasks such as code review, fixing known issues, finding vulnerabilities in source code you own, and initial triage of security alerts remain available. Not all of these general development and maintenance tasks have come to require special verification. Access controls for advanced cyber capabilities are distinguished from the usage conditions for ordinary code analysis and fixes.
The existence of the CVP, however, does not mean exceptions are allowed for restrictions on game production. It also has not been confirmed whether the prompt change to avoid recreating games and the access controls and blocking decisions for cybersecurity use are managed by the same mechanism. The fact that Defense Access permits defensive malware analysis cannot be taken as grounds to interpret copying or porting games as allowed under the same conditions. The two differ both in the work involved and in their institutional position.
Original games remain possible; the accuracy of refusal decisions is the next focus
Osmani said Anthropic will continue to make adjustments to balance protecting others' intellectual property rights against not obstructing legitimate game development. He also asked that anyone refused by Claude Code while developing their own work send in specific examples. The policy of allowing original works is clear, but whether actual responses match that intent is still at a stage that needs improvement.
For developers, what matters is to clarify what is being built and how the rights are arranged, and to record concretely the circumstances in which a refusal occurs. The nature of a request differs depending on whether it is an original work, a work that draws on an existing genre, or a port or recreation of a specific game. For contract work, it also matters what the client owns and on what authority or permission the materials are used. When reporting a refusal, including the model used, the Claude Code version, and the operation that triggered the refusal makes it easier to separate the development content from the conditions under which it occurred.
That said, there is no guarantee that explaining such information will always lift a restriction. Even when the technical content of game development is the same, the decision may change depending on the rights relationships and the source of the materials. How accurately this change can distinguish recreating existing works from original production will need to be checked against future user reports and Anthropic's adjustments.
This change to Claude Code does not prohibit AI-assisted game development itself; it is a measure to avoid recreating other parties' specific works. At the same time, the refusal reports on contract work highlight how difficult it is for AI to properly judge users' rights relationships and creative intent. The key question going forward is how far Anthropic can build a system that prevents intellectual property infringement without needlessly obstructing original works and legitimate development work.