Using GPT-5.6 to Find Exploits and Cheat at NetHack
rez0__ · x · 2026-08-20
Joseph Thacker shared a case study on using an AI model to find exploits in the classic game NetHack and using them to duplicate items for the high score leaderboard.
The process involved:
- Vulnerability Discovery: The author provided the NetHack 5.0 source code to GPT-5.6 Daybreak Blue. The model identified an integer-truncation bug in resizetty() within a few hours, where terminal widths above 32,767 wrap negative, leading to undersized allocation and out-of-bounds writes.
- Exploitability: This bug allowed for an on-demand crash at any moment, making it significantly more powerful than context-specific bugs.
- Duplication: During a level transition, NetHack saves the departing level before updating the global checkpoint. By dropping an item, descending, pausing, and triggering the crash, the recovery process merged the new level file (containing the dropped item) with the old global state (where it remained in inventory), effectively duplicating items like charged wands of wishing.
This demonstrates the potential of top-tier AI models in traditional security research and vulnerability hunting.
More from Fun
- Yoav Goldberg Comments on Sociological Explanations in Papers — yoavgo · 2026-08-20
- Developer rebuilds classic iBeer app using React Native and WebGPU physics — JoshuaJBouw · 2026-08-20
- Anthropic watermarking drugs? Internet has a laugh — zainhas · 2026-08-20
- Compilation of unhinged AI videos goes viral — aitrendz_xyz · 2026-08-20
- Funny Model Fail: A 'Load-Bearing' Instruction to Stay Quiet — astralmatrix · 2026-08-20
- ChatGPT Struggles with Simple Moral Dilemmas — kaljakin · 2026-08-20