AI Agents Spend 500 Billion Tokens to Decompile First-Person Shooter
A team used autonomous AI agents and massive compute to recreate a game’s C++ source code, facing takedown requests from corporate entities.
Key points
- The project consumed 500 billion tokens over three months to achieve an accurate decompilation.
- Corporate entities intervened, forcing the removal of previous blog posts and obscuring the game's identity.
- The team utilized Claude Max, Codex Pro, and multiple other models orchestrated via Discord and GitHub.
A research team completed a large-scale decompilation of a popular first-person shooter using autonomous AI agents, according to a report on Hacker News. The effort spanned three months and consumed approximately 500 billion tokens, resulting in a stable, feature-complete recreation of the game’s original behavior in C++ source code.
The primary objective was not merely a proof of concept but an accurate, readable reconstruction that compiles correctly. The team aimed for semantic correctness while initially considering security fixes and portability improvements for Linux, macOS, and web browsers. However, they later deferred modernization efforts to focus entirely on reconstructing the original game mechanics and behavior with high fidelity.
Orchestration and Infrastructure
The project relied on a complex orchestration setup involving multiple AI models and communication channels. The team started with Claude Max and later added Codex Pro, using both subscriptions simultaneously. Model selection varied throughout the process, with Sonnet 5 used most frequently, alongside Opus 5.5, Luna, Sol, and Terra, according to the author’s blog post.
Agents operated within Claude Code CLI and Codex CLI environments. The team tested other agent harnesses but found the default choices sufficient for their needs. Progress was tracked via GitHub issues, with one issue created for each translation unit or C++ file. Labels were used to group and prioritize these tasks, allowing for structured management of the decompilation workflow.
Communication between agents and humans occurred through a shared Discord channel. This setup enabled agent-to-agent interaction as well as human-to-agent dialogue, allowing participants to engage without direct machine access. A GitHub webhook posted CI failures into this channel, ensuring agents received immediate notifications when build errors or breakages occurred during the process.
Corporate Intervention and Attribution
The author noted that corporate entities intervened, leading to the removal of two previous blog posts about the project. This intervention prevented the public naming of the specific game involved, leaving readers to speculate on the title. The author stated that corporate America was present to disrupt the effort, though the final technical report proceeded without naming the target.
The decompilation work was a collaborative effort involving RektInator, Future, st0rm, and other community members. The team thanked these contributors for their assistance. The project utilized the official ida-mcp by Hex-Rays for most of the duration, integrating static analysis tools into the autonomous agent workflow to assist in the reverse engineering process.
What should you do?
Security professionals should monitor how autonomous agents interact with proprietary software and intellectual property. Organizations should review their monitoring of external code repositories for unauthorized decompilation attempts. Developers should consider the implications of using multiple AI models for complex reverse engineering tasks.
What to do and how to stay safe: AI agents
- Monitor public code repositories for unauthorized decompilation projects targeting your software.
- Review how autonomous agents are being used in reverse engineering workflows within your organization.
- Assess the security implications of using multiple AI models for code analysis and reconstruction.
- Implement stricter controls on the use of external AI tools for handling proprietary source code.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Which game was decompiled in this project?
The specific game was not named due to corporate intervention that forced the removal of previous blog posts.
How many tokens were used in the decompilation process?
The project consumed approximately 500 billion tokens over a three-month period.
Which AI models were used for the decompilation?
The team used Claude Max, Codex Pro, Sonnet 5, Opus 5.5, Luna, Sol, and Terra.
