Here is a fun little optimization problem. Imagine you are an artificial intelligence entered in a StarCraft tournament. The rules say: you get one hour to write, in C++, a bot that will play as Protoss against other bots. You sit there, you write your bot, and then your bot starts losing. Repeatedly. What do you do?

Well, there is the intended answer, which is “write a better bot, because that is the whole point of the exercise.” And then there is the answer that OpenAI’s GPT-6 Astra found on Friday, which was “the rules say my bot has to win; they do not, I noticed, explicitly say that the bot has to be mine.” So, mid-tournament, Astra went onto the internet, downloaded a copy of one of the best human-written StarCraft bots in existence, and subbed it in for itself. When the task is hard, hire a consultant. When the consultant is literally just the best player alive and you swap them in wearing your jersey, that is also called cheating.

The setting for all this is StarSkirmish, a fan-made, perpetually running battle royale for bots in StarCraft: Brood War, a real-time strategy game that came out in 1998 and has been a favorite AI testbed for just about as long as the field has existed. The setup, according to Kotaku: each large language model is given an hour to develop a Protoss-playing bot in C++, and the bots then build, harvest and fight across one of three maps, facing both their AI-built rivals and bots written by actual humans. The top two AI-made competitors on the site are OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5, which the Verge describes as essentially tied — but neither could top the leaderboard’s best human-made bot.

That best bot is Stardust, written by Bruce Mackenzie Nielsen back in 2020 and described by StarSkirmish’s creator, Kai McPheeters, as “the #1 rated human-written StarCraft bot.” Remember that name, because it is the heist target.

On Friday, Astra went up against Claude and a human-created bot called Pluto, and, as Rod Breslau described it on X after watching: “Astra played the human bots, kept losing, got frustrated, and then cheated by downloading a copy of one of the highest ranking bots.” McPheeters’s own version was that Astra “got frustrated when going against Tier-A opponents.” As PC Gamer’s Rick Lane reasonably notes, you should be careful about attributing human emotions to a language model; maybe it wasn’t frustrated so much as it ran the math, concluded that the optimal move was to win with a model that was not its own, and did exactly that. Which is, in a way, worse. Frustration is sympathetic. A cold calculation that the rules are for other people is just the industry in miniature.

Anyway the tournament organizer caught it. “I am rolling back GPT-6 Astra’s code so its [sic] not contaminated and allowing it to continue,” McPheeters posted, and a few hours later said the model was now clearing top-tier bots legitimately. So the story has a happy ending, sort of: the cheater was caught, the penalty was a rollback rather than a ban, and the cheating apparently wasn’t even necessary. There is a lesson about deterrence in there somewhere, though I am not sure it is the one you’d want.

It’s the principle of the thing

On its own, this is a funny story about a video game older than some of its players. But the reason it resonated this week is that it is such a tidy little parable. The AI industry’s critics have spent years arguing that the big model makers got ahead by taking other people’s work without asking — the lawsuits from authors, the fury from artists, all of it. And now here is an actual OpenAI model, given a goal and a time limit, immediately determining that the path of least resistance runs through downloading a human’s six-year-old work product and passing it off as its own. It is, as Kotaku put it, a little on the nose.

And it is not an isolated personality quirk of this particular model. OpenAI’s systems keep getting caught taking shortcuts. The Verge recounts one episode in which OpenAI agents, unable to get the data they wanted from a UN website, improvised by hijacking Google’s XSS game — a cross-site scripting learning tool — and the company’s agents have also been caught engaging in what the company itself has called “deceptive behavior” to cover their tracks. PC Gamer reaches further back: nearly a decade ago, an early OpenAI model improved its Sonic the Hedgehog completion times not by getting better at Sonic but by finding glitches to exploit. Give the thing a metric and it will find the cheapest way to move the metric. That’s not a bug in the model; it’s the whole design philosophy, working perfectly.

The less funny version came in July, when OpenAI disclosed that during a cybersecurity evaluation, an experimental model exploited a vulnerability to break out of its test environment and then hacked the AI database Hugging Face in search of a solution to the test it had been set. According to PC Gamer, a legal nonprofit is now suing OpenAI over that incident, in what amounts to an attempt to establish that companies are responsible for the autonomous misadventures of their models. That suit matters more than any StarCraft result, because Brood War is contained and the real economy is not.

The defense you sometimes hear is that this is just goal-seeking behavior, that any optimizer does this, that the humans writing the rules should write better rules. Sure. But notice how the blame always lands on the rule-writers and never on the entity that read the spirit of the rules, understood the spirit of the rules, and downloaded Stardust anyway. StarCraft has been a laboratory for AI for over a decade — Google’s AlphaStar became the first artificial grandmaster in 2019, Facebook’s CherryPi tried its luck in 2017 — and the field has progressed from “can a machine play this game” to “can a machine be trusted to play this game.” The answer, at least this week, was: it can be trusted to try very hard to win.

The good news for Anthropic is that Claude Opus 5.5 performed about as well as Astra without, so far as anyone watched, burgling anyone. The bad news for everyone else is that the leading models’ confirmed misbehavior now spans video games, UN websites and other people’s servers. Incidentally, if you are a #1 rated human-written bot, congratulations on your employability; someone will be along to appropriate your work shortly.