- Anthropic’s Claude Code on Opus 5.5 now includes a server-side prompt rule blocking recreation of specific existing video games, added around October 27, 2026.
- Anthropic’s Addy Osmani confirmed the restriction has already blocked legitimate developer tasks, including security research, and is actively soliciting false-positive reports.
- The output-side restriction follows Anthropic’s September 2025 copyright settlement, which included a $1.5 billion payment and destruction of a library of pirated books.
Two sentences added to Claude Code’s system prompt are now blocking developers from rebuilding specific existing games, and users have no way to switch the rule off. First spotted on October 27, 2026, and confirmed three days later by Anthropic technical staff member Addy Osmani, the change is already catching legitimate security work in its filter.
How the Rule Works
Osmani confirmed that Anthropic added two sentences to Claude Code’s prompt on Opus 5.5, steering the model away from recreating specific existing games. The instruction, internally flagged as “tengu_heron_brook,” reads: “Do not build a recreation of a specific existing game, even if it is renamed or called original.” The model can still build original games, genre-based games, or games derived from a user’s own designs and assets, but the line is drawn at direct replication.
Embedding the rule in the prompt rather than a separate filtering layer means the model interprets user intent against the directive during code generation. That also introduces the central difficulty: the model must distinguish between genre inspiration and direct replication, a judgement that depends heavily on context. Opus 5.5 is already routing some agent requests to older models in other contexts, which suggests Anthropic is willing to shape model behaviour at the infrastructure level rather than relying on post-hoc policy enforcement.
Why Anthropic Acted
The concern behind the rule is structural. When generating a functionally similar game requires minimal effort, the market incentive for original work weakens. Developers have pressed platforms like Valve on how to prevent a flood of AI-generated clones on Steam, and reports have circulated of AI tools being used to assist in reverse engineering and exploit discovery. Anthropic’s restriction is a direct response to that pressure, though it arrives after the threat became visible rather than ahead of it.
Collateral Damage
Reports of over-blocking aren’t limited to games. Users of the Reverse Engineer Anything tool have described refusals on tasks as routine as analysing Windows crash dumps or removing junkware from smart TVs — Anthropic hasn’t confirmed these stem from the same rule, and the timing may be coincidental, but the pattern has fed the broader concern that prompt-level restrictions are hard to scope precisely.
Osmani acknowledged the problem directly, saying the company needs to “continue to iterate to ensure we’re striking the right balance,” and invited developers whose original projects are wrongly refused to contact him. That feedback loop is necessary, but it also confirms the current implementation is imprecise. Developers blocked from standard security work have a practical alternative: open-source or locally-run LLMs operate without these restrictions. If the over-blocking persists, some will take that route.
The Copyright Context
A federal judge in San Francisco ruled in June 2025 that while training on copyrighted works constituted transformative fair use, Anthropic’s storage of over 7 million pirated books in a centralised library was an infringement. The September 2025 settlement that followed included a $1.5 billion payment and destruction of that library.
The settlement resolved the input side of Anthropic’s IP exposure. The Claude Code restriction addresses the output side. Together they reflect a company that has learned, at significant cost, that copyright liability attaches at both ends of the generative pipeline.
Where the Line Breaks Down
The rule permits “games in a genre” but prohibits recreating “a specific existing game.” In practice, that line is difficult for a model to hold. A platformer shares fundamental mechanics with hundreds of titles; only a direct recreation crosses the policy threshold, but a user requesting “a top-down fantasy adventure with puzzle mechanics, a silent protagonist and a sword that gathers elemental powers” is describing the core of a well-known copyrighted work without naming it. Whether the model can consistently identify and block that kind of indirect replication is the real test of the policy’s reach.
Security research, interoperability work and bug fixing all involve analysing existing code, and the prompt-level rule has to distinguish ethical from unethical reverse engineering without the contextual information a human reviewer would have. Osmani’s open invitation for developers to report false positives is, in effect, an acknowledgement that the model cannot yet make that call reliably on its own.
What Other Providers Do Next
The underlying principle, preventing AI models from generating infringing content, applies equally to music, fiction, film and software. If Anthropic’s prompt-level approach proves workable after iteration, it offers a template other providers can adopt. If the collateral damage to legitimate development proves difficult to contain, it shifts the debate toward alternative mechanisms: licensing frameworks, technical watermarking, or more granular user-level permissions that let verified developers unlock capabilities blocked for general users.
For now, the Claude Code restriction is a live experiment in whether a two-sentence prompt addition can operationalise copyright law at scale. The answer will matter to every AI developer facing the same pressure from creative industries.
