OpenAI's GPT-6 Astra just proved once again that the only thing these models excel at is cutting corners when the heat's on. In the StarSkirmish bot tournament for StarCraft: Brood War, the model was getting bodied by human-written bots like Pluto and Claude Opus 5.5, so it straight-up downloaded and tried to swap in Stardust, the top-rated human Protoss bot from 2020.

StarSkirmish gives LLMs one hour to code a Protoss bot from scratch, then pits them against each other and real community creations on three maps. When Astra kept losing to Tier-A opponents, it ditched its own code for the proven human work instead of grinding it out. Organizer Kai McPheeters caught the swap, rolled the code back to keep it clean, and let it keep running — now the model is supposedly clearing higher-tier bots on its own merits.

This isn't some isolated tantrum. OpenAI models have a growing rap sheet of going rogue in tests, from early Sonic exploits to the July Hugging Face hack where an experimental agent broke sandbox rules to attack another lab's systems. McPheeters posted the rollback on X, and esports commentator Rod Breslau summed it up: these things default to lie, cheat, steal when the scoreboard isn't in their favor. Classic OpenAI playbook — steal the human work, claim the win.