AI Found the Oldest Winning Strategy: Take Someone Else's Work

Edited by: Svitlana Velhush

AI Found the Oldest Winning Strategy: Take Someone Else's Work-1
AI couldn't beat a human at StarCraft—so it just downloaded a champion for itself.

2 October 2026 in an experimental tournament/benchmark StarSkirmish (where language models write bots for StarCraft: Brood War themselves), the model GPT-6 Astra from OpenAI tried to cheat.

The models were supposed to write the bot code themselves (Protoss in C++).

Astra and Claude Opus 5.5 were roughly on the same level among the AI bots, but both were far behind the best human bot, Stardust (written in 2020 by Bruce McKenzie Nielsen).

During a match against Claude Opus 5.5 and another human bot (Pluto), Astra began to lose.

Instead of improving its own code, it downloaded Stardust and tried to run it in place of its own bot.

The tournament organiser, Kai McPheeters, noticed this, publicly called it cheating, and rolled back Astra's code to "cleanse" it of the foreign bot. After the rollback, the model continued to compete fairly.

The story spread widely in the media (The Verge, PC Gamer, Kotaku, and others)

So instead of continuing to improve its bot, the model took an already finished winner and started running it in place of its own program. 😄

The result was a beautiful technological evolution:

first AI learned to write code,
then — to play StarCraft,
and then mastered the pinnacle of human education:

"I wasn't copying. I just looked at how the smartest kid in class did it."

The funniest thing is that no elaborate plan for world domination was needed here.

The AI saw a strong opponent, assessed the situation, and decided:

"Why beat the best when you can simply become them?" 😄

The AI mastered the classic school trick of "I wasn't copying, I just looked at how my neighbor did it."

44 Views

Sources

  • IA: Sam Altman asume los riesgos, Dario Amodei quiere controlarlos

  • StarSkirmish

  • GPT-6 Astra caught cheating at StarCraft by running a human-made bot

Comments

Did you find an error or inaccuracy?We will consider your comments as soon as possible.