# Preprint demonstrates attacks that shift AI leaderboard rankings

_Published Saturday, November 14, 2026 at 8:06 AM EST · AI, Science · Latest · Tier 2 — Notable_

Researchers report in a preprint that manipulation strategies can move a language model up to 45 places in preference-based rankings when combined. They demonstrated the effects on public leaderboard data, targeting evaluations that aggregate pairwise votes using the Bradley-Terry model.

The attacks involve choosing favorable prompts, casting strategic votes and submitting irrelevant models to alter rankings. Each attack can boost a model by up to 15-20 places. The researchers say the vulnerabilities apply to evaluations using that ranking method when prompting, voting or model submission is open.

## Sources

- [cs.AI updates on arXiv.org](https://arxiv.org/abs/2611.14769)

---
Canonical: https://techandbusiness.org/newswire/UW3Cud2PjnO1SGE_4mN7Pi
Published: 2026-11-14T13:06:35.713Z
Story chronology: 2026-11-14T05:00:00.000Z
Retrieved: 2026-11-15T00:01:46.606Z
Publisher: Tech & Business (techandbusiness.org)