Utility-Hijacking Attacks on Instruction-Tuning Data Selection

# Utility-Hijacking Attacks on Instruction-Tuning Data Selection

https://gcdradar.vercel.app/incidents/utility-hijacking-attacks-on-instruction-tuning-data-selecti · AIR-2026-1463 · closed 2026-11-05

- Lane: Research
- Severity: 1 (Research / proof of concept)
- Fix status: Fix status unknown
- First reported: 2026-10-22
- AI role: Attack on AI

## Summary

This research paper introduces utility-hijacking, a systematic supplier-side attack against instruction-data selectors used for fine-tuning. The study shows how harmful records can bypass filters to compromise the safety of fine-tuned models.

## What to do

No action recorded; read the sources below and the vendor's advisory.

## Latest updates

- 2026-10-22: Characterizing and Defending Utility-Hijacking in Instruction-Tuning Data Selection (https://arxiv.org/abs/2610.29886)

## Sources

- 2026-10-22 arXiv: Characterizing and Defending Utility-Hijacking in Instruction-Tuning Data Selection https://arxiv.org/abs/2610.29886

Verify against the vendor's advisory before changing production. Data CC BY-SA 4.0, https://gcdradar.vercel.app.