The BaseResearcharXiv

Characterizing and Defending Utility-Hijacking in Instruction-Tuning Data Selection

According to the abstract excerpt, the authors present a first systematic study of attacks by data suppliers on instruction-data selectors, in which harmful training records are shaped to score well so that selection concentrates the poison. They test three payloads, six selectors and several Qwen victim models.

Open the source ↗

Why it matters

The same thing, for different readers.

For lawmakers

Safety of a fine-tuned model depends on its data supply chain, which the authors say has been studied little.

For researchers

Selectors meant to filter harmful fine-tuning data can be gamed so that top-ranked picks concentrate the poison, according to the authors.

How the finding was reached

In our words, from the source. Weigh the finding by how it was made.

What was done
Adaptive supplier-side attacks using a generate-and-rerank procedure were run against six instruction-data selectors with three payloads (over-refusal, false-fact injection, insecure code) on victim models from Qwen2.5-7B to Qwen3-32B.
What it found
The numerical results were cut off in the excerpt.
What it does not show
Only the truncated abstract excerpt was read; the defence and its limits were not seen.
Status
A preprint: not yet peer reviewed

The record

Questions
Are AI companies putting stronger safeguards into practice? · How is AI used in attacks, defence and war?
People
Yahao Ren (Wrote it) · Jian Lou (Wrote it) · Huadi Zheng (Wrote it) · Jiawei Duan (Wrote it)
What we read
We have read only the description the publisher gives. No quotations until someone reads the source itself.
Found by
The Radar, run of October 22, 2026, via arxiv-ai-safety-query (tier 1: the source’s own channel or an official record)
Cite this
arXiv, “Characterizing and Defending Utility-Hijacking in Instruction-Tuning Data Selection”, October 21, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/arxiv-2026-10-21-utility-hijacking-instruction-data-selection
Our notesLight lane: excerpt cut off mid-result; read the paper before citing numbers. No source text saved.