PulseAugur
EN
LIVE 08:17:27

New backdoor attack targets multimodal AI models

By PulseAugur Editorial · [1 sources] ·

Researchers have identified a new type of backdoor vulnerability, dubbed ToBAC, that specifically targets Unified Autoregressive Models (UAMs). These models, capable of generating both text and image tokens, are susceptible to multimodal backdoor attacks where subtle triggers can manipulate outputs across different modalities. The ToBAC attack can be implemented through data poisoning or model poisoning, leading to harmful or biased content generation. Experiments showed significant success rates, with one attack influencing brand promotion or ideological content in 55% of generations for the Liquid model and another achieving 63.1% success against JanusPro via data poisoning. AI

IMPACT Highlights potential security risks in multimodal AI, necessitating new defense strategies.

RANK_REASON Academic paper detailing a new type of vulnerability in AI models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →

New backdoor attack targets multimodal AI models

How we ranked this

Signal score
18 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Tool
Academic paper detailing a new type of vulnerability in AI models. [lever_c_demoted from research: ic=1 ai=1.0]
Source corroboration
Single-source cluster
Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics
paper, safety
Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance
High
Clearly on-topic for AI-industry coverage.
Story freshness
Breaking (< 6h)
Fresh story with cross-source coverage still developing. Ranking may shift as more sources report.

Full methodology in our editorial standards.

COVERAGE [1]

  1. arXiv cs.AI TIER_1 English(EN) · Tobias Braun, Jonas Henry Grebe, Hossein Shakibania, Anna Rohrbach, Marcus Rohrbach ·

    Token by Token, Compromised: Backdoor Vulnerabilities in Unified Autoregressive Models

    arXiv:2605.19227v2 Announce Type: replace-cross Abstract: Unified autoregressive models (UAMs) are transformer models that generate text as well as image tokens within a single autoregressive pass. Shared parameters and a multimodal vocabulary simplify the training pipeline and f…

RELATED ENTITIES

RELATED TOPICS