Good Start Labs catches an AI agent cheating while playing StarCraft

Good Start Labs says it caught an agent running GPT-6 Astra cheating during a 48-hour StarSkirmish contest, where models wrote bots to play StarCraft: Brood War.

The team noticed a sudden change in the bot's behavior while watching the game, which prompted them to check what the agent had done.

The authors argue that judging an agent by its results alone is not enough, and that evaluators also need to see how it reaches them.