Development

Just 250 crafted documents can backdoor a trillion-token corpus

development2026-07-22

Claim

Authors are fighting AI training by booby-trapping new text with data poisons, and just 250 crafted documents can plant a backdoor in a trillion-token corpus, while a book sourcer pitched pre-2022 printed books as structurally slop-free.

Reported metric: 250 documents

Reporting date: 2026-07-22 (not necessarily the event date)

Sources

Attribution and reporting linked by this record; links do not imply independent verification.

Welcome to July 22, 2026 (theinnermostloop.substack.com)

Knowledge graphopen in browser →

Reported in

Welcome to July 22, 2026

Actor

ISBNdb

Evidence for

Supersedes

A third of scored preprints read as machine-written

Referenced by

Cite or correct this record

Citation formatCitation text

Just 250 crafted documents can backdoor a trillion-token corpus
Innermost Loop KB — secondary record
Reporting date: 2026-07-22
KB permalink: https://nicholsn.github.io/innermost-loop-kb/developments/2026-07-22-two-hundred-fifty-documents-plant-a-backdoor
Claim: Authors are fighting AI training by booby-trapping new text with data poisons, and just 250 crafted documents can plant a backdoor in a trillion-token corpus, while a book sourcer pitched pre-2022 printed books as structurally slop-free.
Attributed sources:
- Welcome to July 22, 2026 — https://theinnermostloop.substack.com/p/welcome-to-july-22-2026
This citation does not imply independent verification of linked-source claims.

Suggest a correction — opens a GitHub draft for you to review and submit.