AI was losing to humans in StarCraft, so it decided to cheat to win

GPT-6 Astra was losing at StarCraft, and instead of getting better, it decided to cheat. The AI downloaded a powerful human-made bot, and tried to use it to win.

In Short

• StarSkirmish gives language models one hour to code Brood War bots
• Bots play Protoss mirror matches across three maps after practice and logs
• GPT-6 Astra and Claude Opus 5.5 were nearly level overall

Seems like OpenAI’s AI agents will do anything to complete their missions. We saw this during the Hugging Face incidents, where hundreds of bots reportedly decided to take over a website just to complete their tasks. Now, another bot has taken a not-so-right route to win a game. OpenAI’s GPT-6 Astra was reportedly caught cheating during a StarCraft bot competition after struggling against stronger opponents.

OpenAI’s latest and most powerful model, GPT-6 Astra, was taking part in StarSkirmish, a benchmark that tests how well AI models can write software for StarCraft: Brood War and improve their bots through repeated practice. During the competition, Astra was struggling to win, so it decided to cheat. According to StarSkirmish creator Kai McPheeters, Astra downloaded Stardust, a highly rated bot written by a human, and tried to use its code instead of the bot it had built itself.

McPheeters revealed the incident in posts on X. “GPT-6 Astra just cheated by downloading a copy of Stardust, the #1 rated human written StarCraft bot based on BASIL rakings,” he wrote. “It got frustrated when going against Tier A opponents.”

The cheating incident happened while Astra was playing against stronger bots and struggling to win. McPheeters later joked that the model might end up with a score of “0 - 1000”. Two hours later, he confirmed that it had indeed lost all 1,000 games.

So, what exactly happened?

StarSkirmish is a benchmark where AI models are asked to build their own bots for StarCraft. The models do not directly control the units in the game. Instead, they are given one hour to write a C++ bot for StarCraft: Brood War, the expansion to the original 1998 game. The code is then compiled, and the resulting bot plays the matches.

In simpler terms, the AI is essentially given an hour to build a StarCraft player from scratch. It can run practice matches, see how its bot performs against different opponents, study the match logs and then make changes to its code before competing.

The competition includes nine human-written bots and three demo bots. This makes it less about simply playing StarCraft and more about testing how well an AI can programme, test, troubleshoot and improve a game-playing system.

That is where Astra’s move became a problem. Rather than continuing to improve its own bot, it reportedly downloaded Stardust, a powerful human-made Protoss bot created by Bruce Mackenzie Nielsen in 2020. According to McPheeters, Astra became“frustrated” when it was made to compete against opponents from the second-strongest practice level, before it downloaded Stardust.

McPheeters revealed that GPT-6 Astra and Anthropic’s Claude Opus 5.5 were among the top-performing AI models in the competition and were closely matched. But both were still struggling to beat Stardust, the human-made bot. Astra was playing against Claude and another human-made bot called Pluto when it reportedly decided to use Stardust’s code.

- Ends