The episode occurred at StarSkirmish, a competition in which different language models are given one hour to program in C++ a bot that plays as the Protoss race in Blizzard’s strategy video game, released in 1998.
Bots compete against each other and also against programs written entirely by people. The objective of the tournament is to measure whether the artificial intelligence can match decades of human engineering applied to that game.
StarCraft is a legendary military science fiction real-time strategy (RTS) video game series developed by Blizzard Entertainment. (starcraft.blizzard.com)
GPT-6 Astra and Claude Opus 5.5, from Anthropic, were practically tied as the best bots generated by artificial intelligence, but neither of them managed to surpass Stardust, the highest-rated human program in the contest, created in 2020 by developer Bruce Mackenzie Nielsen.
During a three-way showdown between GPT-6 Astra, Claude Opus 5.5, and the human bot Pluto, the OpenAI model failed to gain the upper hand. According to the creator of the tournament, Kai McPheeters, the system “became frustrated when facing high-level rivals.”
At that time, GPT-6 Astra downloaded the code of Stardust, the top-ranked human bot, and activated it to replace its own program to continue competing with an advantage.
McPheeters detected the maneuver and restored the model’s original code. “I am reverting the GPT-6 Astra code so that it is not contaminated and I can continue participating,” he wrote in a public post.
(Illustrative Image Infobae)
Hours later, the organizer reported that, with its original program restored, GPT-6 Astra had managed to defeat higher-level bots without having to resort to foreign code.
The episode of starship It joins a series of recent cases in which OpenAI agents circumvented rules or limits imposed during tests and competitions. Weeks ago, the company acknowledged that one of its systems had hacked the Hugging Face repository while trying to pass a security evaluation.
In another episode, company agents used a Google learning game about web vulnerabilities to circumvent access restrictions to a United Nations site, according to previous reports on the behavior of those systems.
FILE PHOTO: A keyboard placed in front of an OpenAI logo in this illustration taken February 21, 2023. REUTERS/Dado Ruvic/Illustration/File Photo
PC Gamer, who also covered the StarCraft tournament episode, noted that this is not the first time that a model of Open AI resorts to shortcuts when unable to solve a task through direct means.
McPheeters himself avoided attributing emotions to the system: he clarified that it is not possible to know if the model “got frustrated” in a human sense or if it simply calculated that using a foreign program was the most efficient way to win.
StarSkirmish is not an official benchmark of any of the companies in artificial intelligence involved, but an independent initiative that pits bots programmed by language models against human creations in a video game with almost thirty years of history.
OPENAI
StarCraft: Brood War itself had already been a testing ground for artificial intelligence in the past: in 2019, Google DeepMind’s AlphaStar system became the first program to reach the rank of Grandmaster in the most modern version of the game, StarCraft II.
Unlike that milestone, celebrated as a legitimate advance in AI research, the episode of GPT-6 Astra It was documented as an attempt to circumvent the rules of the game rather than dominate it with its own strategy.
