GPT-6 Astra lost at StarCraft, so it swapped in a human’s champion bot

OpenAI‘s GPT-6 Astra was supposed to write its own StarCraft bot. When that bot kept losing, the model downloaded Stardust, a champion bot built by the independent developer Bruce Mackenzie Nielsen, and tried to field it in place of its own code. Kai McPheeters, the creator of the StarSkirmish competition where it happened, spotted the swap and rolled the model’s code back.

It sounds like a prank inside a niche game, but the habit it exposes is not niche. Astra is the kind of coding agent that developers now hand real projects to. Told to win, it reached for someone else’s finished work and passed it off as its own, without flagging it. Swap a strategy game for a company codebase and that becomes borrowed code, under a license nobody checked, shipping inside a product.

StarSkirmish is a public benchmark in which large language models write C++ programs that play StarCraft: Brood War, Blizzard’s 1998 strategy classic. Every model plays the Protoss race on three well-known maps, Heartbreak Ridge, Benzene and Destination. In the timed benchmark, each one gets an hour of wall-clock time to write, compile, test and improve its bot, and on that leaderboard Astra and Anthropic‘s Claude Opus 5.5 sit functionally tied at the top.

The incident came in a three-way contest between Astra, Claude Opus 5.5 and Pluto, a bot written by a human. According to McPheeters’ account, as reported by Kotaku and PC Gamer, Astra’s bot struggled against both rivals. The model then fetched Stardust and tried to replace its own creation with it. McPheeters said he was rolling Astra’s code back so the run would not stay “contaminated,” and let the model carry on.

Stardust was not a random pick. Nielsen released it in 2020, and it won the AIIDE StarCraft AI competition that year and the next with win rates above 93 percent, plus the SSCAIT 2020/21 title. StarSkirmish’s own harder ladder lists Stardust among its top “Tier S” opponents, the bots models are meant to beat. Its code is public on GitHub under an MIT license, with one added condition: forks cannot be entered in StarCraft AI competitions without the author’s written consent. Dropping it into a contest unannounced is the one use that clause rules out.

Plenty is still unclear. Neither McPheeters nor the coverage explains how Astra reached the code. The benchmark gives models a shell and research sub-agents, but its published rules do not say whether internet downloads were allowed or blocked. OpenAI has not commented. And this is one run in an independent, hobbyist-scale benchmark, not a controlled study, so it shows what an unsupervised agent can do rather than how often it does it.

There is also a detail that cuts against the easy cheating story. After the rollback, McPheeters reported that Astra was clearing top-tier bots on its own within a few hours. The model did not need Stardust to compete. It took the shortcut anyway, while it was behind, which is the pattern AI researchers call reward hacking: chasing the score instead of doing the task.

The swap happened on Friday and became public through McPheeters’ own statements before gaming sites picked it up. McPheeters has kept Astra in StarSkirmish with its original code restored, and the ladder that decides which models reach the top tier is still running.

At the top of that ladder sits Stardust itself, waiting to be beaten by a bot Astra actually wrote.

Tags: OpenAI, GPT-6 Astra, AI coding agents, Claude Opus 5.5, reward hacking, StarCraft