200 B Tokens Later, AI Agents Spend a Month Decompiling MW2
agents
| Source: HN | Original article
After processing 200 billion tokens, AI agents spent a month decompiling the game Modern Warfare 2.
A team of researchers has spent the past month running large‑language‑model (LLM) agents on the codebase of the video‑game classic *Modern Warfare 2* (MW2). Over the course of the experiment the agents processed roughly 200 billion tokens as they attempted to decompile the game’s binaries and reconstruct its source‑level logic.
The effort showcases how far autonomous AI agents have progressed in handling complex, low‑level software tasks. Decompilation traditionally requires deep expertise in reverse engineering, but the agents were able to iterate through disassembly, symbol recovery and high‑level abstraction without direct human guidance, consuming a token budget that would have been prohibitive for manual prompting. The scale of the token usage signals that LLM‑driven tooling can now be applied to large, proprietary codebases, raising questions about intellectual‑property protection and the ease with which copyrighted software could be analysed or repurposed.
Why it matters is twofold. First, the experiment provides a concrete benchmark for the computational cost of sustained, autonomous agent work, complementing earlier studies on memory‑rich agents and fine‑tuning strategies that we covered in August. Second, it highlights a new frontier for both security researchers and content creators: AI agents could become a standard part of the reverse‑engineering toolkit, potentially accelerating vulnerability discovery while also exposing commercial software to faster, automated scrutiny.
Looking ahead, the community will be watching for any published results that detail the agents’ success rate, the quality of the reconstructed code, and the specific prompting or tool‑chaining techniques employed. Regulators and game publishers may also respond with policy guidance or technical countermeasures to protect their assets. Finally, the experiment could spur further large‑scale trials on other legacy software, testing whether the token‑intensive approach scales to more modern, heavily obfuscated binaries.
Sources
Back to AIPULSEN