They Found Their Own Code in Google's ARTEMIS — and a Force-Push That Erased Their Names

The sentence that will likely define this story was not written by a Google engineer. It was written by Nicolas Dehandschoewercker, CEO of Paris-based mobile-agent startup Minitap, in a blog post published September 11, 2026: “We opened Google’s Artemis repository and recognised code we had written for mobile-use. Then we found our names in its history — and the change that removed them.”

Google released ARTEMIS earlier this month as an open-source framework for turning natural-language instructions into reliable Android automation. It integrates with AI coding assistants such as Antigravity, Codex, and Claude Code, and it claims a 99%+ success rate on the AndroidWorld benchmark. The repo, published under the google GitHub organization and built by the Pixel Test Engineering team, picked up 3,400 stars and glowing coverage within days. Then Minitap’s engineers opened it and, in the words of their CEO, said: “What the fuck? We wrote this.”

What Minitap says it found

Minitap built mobile-use as an open-source research experiment — code for letting AI agents reliably interact with phones — and published it under Apache 2.0. The team worked on it through February 2026, then moved on to a newer, closed-source version that now powers their commercial product. The public repo stayed up, as open-source code does.

According to Minitap’s post, the overlap with ARTEMIS is not generic or coincidental. It is specific, and in places identical:

• Device-connection code. Parts of the code for connecting to Android devices “matched our implementation exactly.”
• The Hopper agent’s prompt. The agent’s system instructions were “identical, word for word.” Hopper, it turns out, is a Minecraft reference — engineer Jean-Pierre Lo liked the name. Seeing the same name attached to the same instructions in Google’s repo “felt very familiar.”
• The Alice/Bob/Charlie example. An example task for keeping an agent inside WhatsApp had the same goal — sending Happy New Year messages to Alice, Bob, and Charlie — with the same comments and cleanup steps.
• A shared bug. Older versions of both projects had a helper that would write a results file, then fail when it tried to read its own output on the next run. Minitap says it reproduced the same failure in both implementations. ARTEMIS has since fixed it.

The most serious claim concerns authorship. An earlier ARTEMIS package file listed three names: Pierre-Louis Favreau, Jean-Pierre Lo, and Nicolas Dehandschoewercker. The version that replaced it removed all three and substituted another author. The only change to the file was the author list. GitHub’s activity record, Minitap says, shows the replacement happened through a force push in August — before their investigation began in September. The README, as of their September 11 check, did not credit mobile-use.

Why the license matters

This is where the story moves from awkward to consequential. Apache 2.0 is one of the most widely used open-source licenses in the world, and it is deliberately permissive — but permissive is not the same as attribution-free. Section 4 of the license requires redistributors to preserve applicable copyright and attribution notices and to identify changes made to the code. Minitap’s post points out that there are ordinary, established ways to do this: a fork links back to the original project, an imported copy documents its source, copyright and attribution notices stay with the code, and a README explains which parts came from elsewhere.

“You should not need to recover an old version of a repository to discover that relationship,” the post argues. That framing is important because it gets at the actual cost of attribution failure. Maintainers already spend unpaid time answering questions, reviewing contributions, and keeping projects working. When a trillion-dollar company republishes their work and they then have to spend days reconstructing git history to prove what they wrote, sharing starts to look like a bad deal. “An ecosystem that makes this routine gives people another reason to stop doing it,” Dehandschoewercker writes.

The leaderboard angle

There is a second, quieter thread to the story: benchmarks. Minitap says it had spent months trying to get its newer AndroidWorld results reflected on the public leaderboard. The maintainer applied their earlier submissions, up to 91.4%, with the last confirmation in December 2025. Minitap then submitted 94.8%, and later 100% in a January evaluation, followed by two follow-up emails — four emails total, no reply. As of September 11, the leaderboard still showed mobile-use at 91.4% while ARTEMIS appeared at 99.1%. Both are self-reported; the leaderboard explicitly says it does not independently verify results.

ARTEMIS’s own comparison chart, meanwhile, omitted mobile-use entirely — while including DroidRun at the same listed score and an unrelated MadeAgents project confusingly called “MobileUse” at a lower score. Minitap is careful to state it has no evidence connecting the unanswered emails or the chart omission to the removal of their names. But taken together, the pattern reads as a small company struggling to get the public record to reflect its work at all — on code and on benchmarks.

Google’s track record cuts both ways

The sting for Minitap is that Google has historically done attribution well. The company gave the world Kubernetes and TensorFlow. When it launched Chrome, it explicitly credited WebKit and Firefox: “We owe a great debt to many open source projects.” Google even publishes its own guidance for handling third-party code. “That is exactly why I expected better,” the post says.

The good news for the ecosystem is that the process appears to be working, however slowly. Google opened a credit-restoring pull request after Minitap raised a public issue on the ARTEMIS repository, and the story climbed to #3 on Hacker News, generating exactly the kind of scrutiny that attribution norms depend on. A public factual record with detailed comparisons, archived files, and a timeline is available for anyone to inspect. Minitap has asked for three things: acknowledgment that ARTEMIS is derived in part from mobile-use, credit for the people behind it, and corrected attribution.

There is also a competitive footnote that reframes the whole episode. The version of mobile-use powering Minitap’s product today is closed source and, in the CEO’s words, “a completely different beast by now.” His parting line to Google: “You’re seven months late.”

Why this matters beyond one repo

Open-source AI is having a complicated moment. Labs open-weight models while keeping training data and methods proprietary; agents that scrape, remix, and republish code are now a normal part of the development loop. In that environment, attribution isn’t bureaucratic pedantry — it is the mechanism that lets the next maintainer decide whether publishing is worth it. Apache 2.0 says you can take the code. It also says you must say where it came from.

If a company with Google’s engineering discipline and open-source pedigree can ship a repo whose README omits a direct upstream dependency — while the upstream’s authors watch their names vanish from the package file via a force push — the lesson for smaller maintainers is bleak. The counterweight is publicity: this story spread because Minitap documented everything, archived the evidence, and published a timeline. Attribution disputes used to be settled quietly, if at all. Now they are settled on Hacker News, with receipts.

The unresolved questions are the ones worth watching. Who at the Pixel Test Engineering team made the decision to replace the author list, and was it reviewed? Will Google respond substantively to the leaderboard submissions, or only to the code attribution? And will ARTEMIS’s README end up looking the way Apache 2.0 always intended — with the original authors’ names restored where they belong?