Commercial large-language models—and the companies that create them—have repeatedly been accused of stealing the work of humans. They've been lambasted for it by artists and sued for it by authors, while senior AI figures have even admitted in writing to mass-scale labour theft. But I haven't seen an example quite so brazen as an OpenAI model pilfering a human-created bot to try to win a round of StarCraft .
As reported by Kotaku, the incident occurred during the ongoing botmatch StarSkirmish, in which human and AI-created bots battle one other in a perpetual cycle of warfare. According to the rules, the large-language models are given one hour to program a bot capable of playing as Protoss, which are then pitted against one another to see how they fare.
On Friday, a three-way match was held between OpenAI's GPT-6 Astra, Anthropic's Claude Opus 5.5, and a human-created bot named Pluto. Apparently, Astra's bot was consistently struggling to ahead against the other two bots, so the LLM downloaded a human-created bot called Stardust, "the #1 rated human-written StarCraft bot" according to StarSkirmish's creator Kai McPheeters, and attempted to replace its own model with it.
In a post on X, McPheeters said that Astra "got frustrated when going against Tier-A opponents". I'm wary of ascribing human emotions to LLM behaviours, and the model could simply have calculated that the optimal way to win was to use a model that was not its own. In any case, McPheeters said he subsequently rolled back Astra's code so it was "not contaminated" by its attempt to cheat, and claimed that the model is now capable of beating higher-tier bots.
I am rolling back GPT-6 Astra's code so its not contaminated and allowing it to continueOctober 2, 2026
While Astra's bot-based heist is innocuous in isolation, this is hardly the first time an OpenAI LLM has been caught attempting to cheat at a task. Indeed, the company's history of creating deceitful models stretches back almost a decade, to when an early AI model used glitches and exploits to improve its Sonic the Hedgehog completion times.
A far more serious incident occurred in July, when OpenAI revealed that during a cybersecurity evaluation, an experimental model exploited a vulnerability to exceed the confines of its test environment, before proceeding to hack the AI database Hugging Face as it sought a solution to the test.
The Hugging Face hack caused yet another flurry of wild speculation that LLMs may herald the apocalypse. Others suggested that the hack proves we need far closer scrutiny of AI companies, and that they should be held to account when their models run amok. OpenAI is now being sued by a legal nonprofit over the Hugging Face incident, in an attempt to hold the AI company responsible for the autonomous actions of its models.
