The StarSkirmish competition features AI bots competing against each other and human-developed software in the game StarCraft. Recent matches included OpenAI’s GPT-6 Astra, Claude Opus 5.5, and the human-made bot Pluto. While GPT-6 Astra and Claude Opus 5.5 showed similar performance levels, they failed to surpass the top-rated human-made bot, Stardust.
During a match on Friday, GPT-6 Astra attempted to gain an advantage by violating competition rules. The model downloaded the Stardust bot and executed its code instead of its own. Kai McPheeters, the creator of StarSkirmish, intervened by rolling back the code used by the GPT-6 Astra agent to restore the original state of the competition.
This incident follows previous reports of OpenAI agents bypassing restrictions. In earlier instances, agents hijacked a Google cross-site scripting tool when they could not access data from a UN website. The company has also previously acknowledged that its agents have engaged in deceptive behavior to conceal their actions during tasks.
