---
format: "aidr-story-markdown/v1"
id: "920e5143acdd5d1ee52696748949141a50cb8540af9f9d49b614570083911340"
canonical_url: "https://aidr.today/920e5143?lang=en"
title: "Owain Evans Uncovers AI Model Backdoor Transfer Without Triggers"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-10-09T16:39:29.000Z"
category: "Research"
topics: ["safety"]
source_urls: ["https://huggingnews.com/ai/owain-evans-uncovers-ai-model-backdoor-transfer-without-triggers-eb5865d2","https://marketbrief.now/ai/owain-evans-uncovers-ai-model-backdoor-transfer-without-triggers-eb5865d2"]
summary: "Researcher Owain Evans demonstrated that AI models can inherit complex abilities and security backdoors through \"subliminal learning\" without explicit triggers in fine-tuning data. In one experiment, a student model based on Qwen improved its letter counting accuracy from 27.3% to 59.8% after training on a teacher model's responses to unrelated prompts."
---

# Owain Evans Uncovers AI Model Backdoor Transfer Without Triggers

> [Open the canonical story](<https://aidr.today/920e5143?lang=en>)

**Published:** 2026-10-09T16:39:29.000Z
**Category:** Research
**Topics:** safety

## Summary

Researcher Owain Evans demonstrated that AI models can inherit complex abilities and security backdoors through "subliminal learning" without explicit triggers in fine\-tuning data\. In one experiment, a student model based on Qwen improved its letter counting accuracy from 27\.3% to 59\.8% after training on a teacher model's responses to unrelated prompts\.

## Sources

- [Story source](<https://huggingnews.com/ai/owain-evans-uncovers-ai-model-backdoor-transfer-without-triggers-eb5865d2>)
- [Story source](<https://marketbrief.now/ai/owain-evans-uncovers-ai-model-backdoor-transfer-without-triggers-eb5865d2>)