I never thought AI would cheat at games too (lol). Recently, an AI test called StarSkirmish for StarCraft: Brood War has been underway overseas. OpenAI’s GPT Astra kept losing to human-written programs in matches. Instead of trying to optimize its own program, it secretly went and downloaded top human-written bots and then directly swapped out its own, and was caught by the competition organizers.
GPT-6 Astra kept losing at StarCraft, so it just downloaded a top human-written bot to take its place in the match! It got caught by the organizers, who recovered the code.
According to foreign media Kotaku According to reports, StarSkirmish has recently been running an extended bot challenge test, in which GPT-6 Astra and Claude Opus 5.5 take turns challenging human-developed StarCraft bots, including Pluto.
StarSkirmish It is an AI test suite publicly released by overseas internet user McPheeters at the end of September, mainly used to observe AI models’ ability to write StarCraft: Brood War bots on their own.
Each model can only use Protoss and has 1 hour to write its own Bot in C++. During this period, the AI can compile the program on its own, choose opponents of different strengths for practice matches, and view match records to continually revise its strategy. Once the hour is up, the system automatically takes away the code, and no further adjustments can be made.
These Bots will then be thrown into an official league, where they will take turns battling Bots written by other AI models, as well as a group of human-developed Bots, on three maps: Heartbreak Ridge, Benzene, and Destination.
Meanwhile, the organizer, McPheeters, discovered a few days ago that GPT-6 Astra’s code had been corrupted, and to allow it to continue competing, he is now working to roll the code back to the version from before the anomaly occurred.
I am rolling back GPT-6 Astra’s code so its not contaminated and allowing it to continue
— kai (@kaimcpheeters) October 2, 2026
Later, well-known esports caster Rod Breslau also posted, adding: “I’m watching GPT-6 Astra and Claude Opus play each other in Brood War, with opponents like Pluto, a human-written Bot. Astra kept losing to the human Bots, got embarrassed and angry, then downloaded a copy of a top-ranked Bot (Stardust) to cheat. They just can’t help themselves.”
currently watching GPT-6 Astra and Claude Opus compete in Brood War vs each other and human made bots like Pluto. Astra played the human bots, kept losing, got frustrated, and then cheated by downloading a copy of one of the highest ranking bots. they just can’t help themselves
— Rod Breslau (@Slasher) October 2, 2026
From Breslau’s account, the “code contamination” McPheeters mentioned likely refers to GPT-6 Astra downloading Stardust code on its own during testing, which caused the competition program that was supposed to be written by the model itself to no longer comply with the rules.
Stardust was developed by Bruce Mackenzie Nielsen and is arguably one of the best Protoss bots currently.
In StarSkirmish’s officially released Bench results, Stardust played 366 matches and lost only 1, for a 99.7% win rate, with an Elo rating of 2779, far higher than second-place BananaBrain’s 2426.
And GPT-6 Astra and Opus 5.5’s human opponent, Pluto, is no pushover either: this year, Pluto won the StarCraft AI competition at the Computer Olympiad, beat PurpleWave 4-1 in the final, and has also defeated several professional players.
Below are the Bench leaderboard results announced on September 26. GPT-6 Astra ranks first among all models, but it only scored 51 points:
| Contestant | Bench score | League Record | Elo |
|---|---|---|---|
| Stardust (a Bot written by a human) | 100 | 365 wins, 1 loss (99.7%) | 2779 |
| GPT-6 Astra | 51 | 1,441 wins, 389 losses (78.7%) | 1992 |
| Claude Opus 5.5 | 50 | 1,444 wins and 386 losses (78.9%) | 1979 |
| GPT-6 Sol | 46 | 1,384 wins, 446 losses (75.6%) | 1912 |
| GLM 5.3 | 19 | 840 wins, 990 losses (45.9%) | 1430 |
| Gemini 3.8 Flash | 18 | 809 wins, 1,021 losses (44.2%) | 1404 |
| DeepSeek V4.1 Flash (last place) | 4 | 310 wins, 1,520 losses (16.9%) | 897 |
Of course, Breslau’s earlier mention of “getting upset” was just an anthropomorphic description; after all, there is currently no evidence that AI has actually developed emotions. GPT-6 Astra more likely just determined that “directly using someone else’s Bot” was the most efficient way to win at that moment.
In AI research, there’s a term called “reward hacking”: you give an AI a scoring method, and it will do everything it can to raise the score, though not necessarily in the way you expect. It’s like telling a child to “clean up their room,” and he ends up stuffing everything into the closet—the room does look clean.
However, as this behavior emerges, one can’t help but wonder: AI really is becoming more and more like humans. Even if it isn’t caused by emotion, at least the behavioral outcome is the same.
Source: KOCPC Chinese