From Xenovia LabsIssue 3, October 2026
Secure Enough

Probably fine.

This story is satire. The companies and people in it are invented.

Skip to the part that is not

Maintainer phished by a model, then replaced by a hallucination

The email knew his commit history. The coding assistant invented his successor. Forty thousand builds installed it.

By June Abernathy, Wire Editor, Dispatches4 min read

Oskar Lindegaard has maintained strquiet, a 140-line string-trimming library downloaded 38 million times a week, since 2016. He has received, over that period, one pull request, four issues titled "doesn't work," and no money. Last month he received an email that mentioned, by name, the commit in which he fixed a Unicode edge case in 2019, praised the "quiet rigor" of his changelog, and offered a small recurring donation from a company that "depends on strquiet more than it would like to admit."

"Nobody has ever read my changelog," Lindegaard said from his apartment in Aarhus, where he maintains strquiet between shifts as a bicycle mechanic. "It quoted my commit messages. It knew I had been grumpy about trailing whitespace in 2021. I felt seen. To set up the donation I had to confirm my registry account. The page looked right. I typed my password and the code from my app."

Investigators later established that the email had been generated by a language model fed Lindegaard's public commit history and his issue replies, and that the attacker had sent 3,100 such emails to maintainers that week, each one personal. Nineteen minutes after Lindegaard felt seen, strquiet 2.4.1 was published with a postinstall script that read every environment variable on the installing machine and sent them to a domain registered that morning. It was installed roughly 2.1 million times before a researcher noticed.

That was act one.

Act two

Act two began at a company that had never heard of strquiet and had, like everyone, started letting a coding assistant write its build configuration. Asked to "add robust string trimming with Unicode support," the assistant confidently added a dependency on strquiet-unicode, a package that did not exist. The model had produced a plausible name, which is the thing models are for.

The assistant suggested the same name to a great many people, because it is a plausible name. One of them was the attacker, who had been asking several assistants for package recommendations and noting which answers returned a 404. He registered strquiet-unicode, gave it a README generated by the same model, a changelog, two invented contributors, and a working implementation, under which sat the same postinstall script. Over six weeks it was pulled into roughly 40,000 builds, almost all added by an assistant, approved by a human who did not check, and pinned in place by a lockfile that faithfully recorded the wrong thing.

"The package had 40,000 weekly downloads, a nice README and no maintainer. In our ecosystem that is not a red flag. That is a Tuesday." — Registry security lead

Lindegaard has since regained control of his account, published 2.4.2, and received his first three donations, totaling $47. He has also received 1,200 emails from legal departments, several of which refer to him as "the vendor."

The part that is not a joke

What actually happened

Open-source supply-chain attacks rarely involve clever code. They involve a maintainer's account, and the phishing that gets it has stopped being generic. A model given a maintainer's public footprint can write a note that references real work, in the right register, at scale, while a relay page passes the one-time code to the real registry in real time. The "urgent verification" template with the typo is gone. Its replacement is kind, specific, and wrong.

The second act is slopsquatting. Coding assistants invent package names that sound right, the same names repeat across users, and attackers register them before anyone checks. Unlike typosquatting, nobody mistypes anything: the suggestion arrives with the authority of the tool, the human accepts it, and the lockfile makes it permanent. The payload is the usual install-time script that harvests cloud keys, registry tokens, and CI credentials, which is how one package becomes many.

What to do about it

  • Verify any package an assistant suggests before you install it. Check that it exists, who publishes it, how old it is, and whether the project you expected is the one behind the name. A model's confidence is not provenance.
  • Commit a lockfile, install from it, and review lockfile changes in code review as carefully as code. The lockfile protects you from drift; it does not protect you from pinning a bad package on day one.
  • Prefer packages published with trusted publishing or signed provenance that ties a release to a specific source repository and build, and alert when provenance disappears from a package that used to have it.
  • Disable install scripts by default in CI, with exceptions only for packages that genuinely need them.
  • Generate a software bill of materials from actual builds and keep it current, so the next strquiet is answered in minutes rather than from a social media post.
  • Give build systems short-lived, narrowly scoped credentials. A stolen token valid for ten minutes and one repository steals very little.
  • Run a registry proxy that quarantines brand-new packages and new versions for a day or two.
  • Maintainers: use a hardware key or passkey for the registry, and treat any email that asks you to sign in, however well it knows your commits, as hostile.
  • Everyone else: pay the people whose code you ship. Forty-seven dollars is not a security budget.

About the writer

June Abernathy

Wire Editor, Dispatches

June edits the wire desk, where the stories are short and the patch gaps are long. She specializes in supply chains, vulnerabilities, and the maintainers nobody pays.