How AI changes games: mods, reverse engineering and mechanics

Parkour from Mirror’s Edge appears in Skyrim, while Minecraft actions are passed into GTA. Similar-looking demonstrations can use different techniques: connecting two running games, reconstructing part of a program or transferring an individual mechanic. Each approach has its own tools and limitations.

The detailed AI Search review of October 6, 2026 compares these approaches. The examples and assessments below come from that recording; tool availability refers to its date.

The review shows Minecraft inside GTA, Spider-Man in a Batman game, Mario in another game, and Mirror’s Edge parkour in Skyrim. The presenter describes these combinations as already working fairly seamlessly and their creation as something users can request from a model without much technical knowledge. Several different approaches can produce such results.

The presenter identifies the release of GPT-6 Astra a few weeks before the recording as the starting point. OpenAI’s GPT-6 Astra page reports SRE-Bench results for reconstructing the core logic of binary programs without source code: 88.0% in one attempt and 99.2% within four attempts. Developers write source code and compile it into a distributable program. Reconstructing its implementation from that result has traditionally been difficult; he believes newer models sharply reduce that difficulty.

He suggests that Claude Opus 5.5 performs even better and claims AI has already fully decompiled games including Super Smash Bros., Mario Kart 64, Halo, Kingdom Hearts and Fire Emblem. He extrapolates from the benchmark to the possibility of reconstructing almost any game or application, naming Photoshop, Microsoft Office and potentially an operating system. These are his assessments of the implications; the quoted percentage concerns the stated benchmark and four attempts.

The first approach is a pass-through mod: the original games remain in place while AI creates an intermediate software layer that transfers actions between them. In the Minecraft/GTA example, placing TNT at a location translates into spawning an explosive object at the corresponding position in GTA. When the TNT detonates, the mod tells GTA to trigger an explosion there. Both games must be running on the computer. The presenter gives Spider-Man inside a Batman game as another example.

He considers this the most limited approach because behavior still depends on the original engines, and interaction between the systems can cause physics and collision problems. He points to Universal Modder and its ready-made mashup skill for creating this layer. Users can give an agent the repository and request a connection between two games. He names GPT-6.1 Sol and GPT-6 Astra as options and relays other people’s favorable reports about Opus 5.5.

An alternative is Reverse Engineer Anything, a repository containing ready-made skills and MCP tools for the agent. MCP provides a standard for connecting an agent to external tools. The next approach involves rebuilding the software that runs the game. The illustration is a vanilla World of Warcraft environment with a character on a skateboard. The model receives the game and uses a reverse-engineering tool such as Ghidra to investigate animations, maps, models and mechanics. It then attempts to rewrite the game, potentially in another language such as Rust.

He attributes the popularity of Rust rewrites to his impression that current frontier models are good at the language; Rust is not presented as mandatory. The described process uses a copy of the game the user owns. Moving to another engine, he says, permits more customization. Another reported example rebuilds Call of Duty: Modern Warfare in Rust and combines it with a Minecraft world, allowing players to shoot mobs, destroy blocks with guns and blow them up with grenades.

Game analysis uses Ghidra and a separate MCP server through which an agent can access it. IDA is another option with its own MCP integration. The review mentions ILSpy for .NET applications and Cpp2IL for Unity games using IL2CPP. The general procedure is to give an agent suitable tools and a copy of the game to analyse, then request a rewrite.

In our project on an agent with permitted actions, this approach serves a catalogue and orders: the agent retrieves data and calls the tools the application provides. Game modification and a commerce service perform different tasks; in both, the connected application defines the available actions.

Decompilation can be followed by a move to another platform. The review includes a Super Mario Galaxy VR mod for Meta Quest and similar work on Mario Kart. It also mentions AnyPS5 and PortPS5, projects using AI to help port PS5 games to PC. The presenter calls God of War: Sons of Sparta and Dead Cells: Return to Castlevania successful examples; that assessment is his.

The third approach extracts an individual mechanic and transfers it to another game. According to the presenter, Mirror’s Edge parkour was added to Skyrim in roughly a day using Claude Opus or GPT Astra. He then lists Skate 3 trick physics in Call of Duty and World of Warcraft, a Minecraft world combining Skate 3 and Call of Duty dynamics, and a combination of Skyrim, Escape from Tarkov and Spider-Man. Here the focus is on transferring particular behavior.

He sees this as a way to extract physics systems, mechanics and animations from existing games for use in a new project without implementing the corresponding component entirely from scratch. He returns to Universal Modder’s mashup skill and Reverse Engineer Anything. The recording also mentions a supplementary guide with a complete pass-through example, a Rust rewrite example, and guidance on prompts and workflow. The guide’s link was unavailable when checked.

The presenter believes no special prompt wording is necessary: current models, especially Opus 5.5, can understand a requested change or character transfer when given access to the tools. He returns to proprietary applications such as Photoshop, Premiere Pro and Microsoft Office as potential targets for the same techniques. Earlier he acknowledged legal implications around cracking Photoshop. His closing remarks about Adobe subscriptions and running its software free are sarcastic; he does not perform that scenario.

At the end, he repeats his broad interpretation of the result above 99% and anticipates major consequences for games and software. He also speculates that model developers may already be reverse engineering games and applications and using the resulting code to train future systems. He supplies no confirmation of that speculation in the recording. He expects subsequent models to improve at creating games and software from descriptions. The question of future legal implications remains open.