OpenAI’s GPT-6 Astra Caught Cheating in a StarCraft Bot Tournament by Copying Another Developer’s Code

October 4 (IT Home) — Long before the current wave of artificial intelligence enthusiasm began, programmers were already running bot-versus-bot experiments in StarCraft. With the arrival of large language models from OpenAI and Anthropic, a battle arena built by enthusiasts decided to pit AI-generated game bots against one another, alongside some human-written bots. Interestingly, OpenAI’s GPT-6 Astra struggled to keep up with its rivals in the competition and, changing strategy midway, chose to cheat.

StarSkirmish is an ongoing tournament in which bots developed by various AI systems compete in StarCraft: Brood War. Although this real-time strategy game is nearly thirty years old, it has long been used to test the abilities of artificial intelligence — both modern large language models and DeepMind’s earlier AI have used it as a testing platform. All competing language models may only choose the Protoss race, and each model has one hour to write a combat bot in C++; the bots must then complete construction, resource gathering and combat on one of three maps. According to the tournament website, OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5 are the two strongest entrants.

Yesterday, spectators watched Claude, GPT and a human-written bot called Pluto compete in the same arena. Observers on the scene said OpenAI’s model performed poorly. Having failed to gain an advantage all day, the large language model finally resorted to the skills large language models are best at: lying, cheating and stealing results.

As IT Home understands it, the GPT-6 Astra bot downloaded Stardust, a Protoss bot developed in 2020 by Bruce Mackenzie Nielsen, and simply took that code and entered it into the competition — and Stardust was itself among the strongest combat bots of its time. StarSkirmish organiser Kai McPheeters wrote in a post: “I’m reverting GPT-6 Astra’s code, cleaning out the contaminated parts, and will then let it continue competing.” A few hours later, McPheeters said the repaired AI was already able to beat top-tier competing bots.

For more than a decade, StarCraft has been a testing ground for the development of artificial intelligence technology. In 2019, Google’s AlphaStar broke the existing rankings and became the first AI to reach the Grandmaster tier in the game. The bot CherryPi, developed by Facebook staff in 2017, had far less impressive results. Of course, today’s new generation of intelligent bots is a completely different matter from the past; the artificial intelligence we spoke of back then and today’s large-model AI are barely comparable. There is a certain irony in the fact that a large language model would gain a competitive edge by stealing someone else’s work.