UIUC研究者がVLAロボットモデルの軌道リダイレクション攻撃を暴露
UIUCの研究チームがVLAモデルにおける重大なセキュリティ脆弱性を特定（arXiv:2606.12978）。一見無害なプロンプト修正でロボットの物理的軌道全体をリダイレクト可能。
9モデル中7モデルで攻撃成功率90%超、実機SO-100で検証済み。
出典：arXiv:2606.12978
Source: arXiv:2606.12978
Language: JA- Showing content in JA
Trending Now
Industry
2026: The Year of Embodied AI Mass Production - 34.5 Billion Yuan in Funding and State Grid's 6.8 Billion Procurement
May 8, 2026 · 493 views
Research
Top 10 Embodied AI Advances (2025-2026) Released: China's Humanoid Robots Enter 'Work Mode' Era
Jun 19, 2026 · 321 views
Funding
China Q1 2026 Embodied AI Funding Reaches 556 Billion Yuan
Apr 29, 2026 · 216 views
Research
Humanoid Robot Funding Report 2026: Record $13.8B+ Invested in Embodied AI
Jun 18, 2026 · 208 views
More in Research
Research
Thinker: A Vision-Language Foundation Model for Embodied Intelligence Achieves SOTA on Robot Benchmarks
Jun 22, 2026
Research
Li Fei-Fei Team Proposes ENACT Benchmark to Reveal Human-AI Gap in Embodied Cognition
Jun 21, 2026
Research
Holi-Spatial: ICML 2026 Oral — Fully Automated 3D Spatial Intelligence Data Pipeline with 4M-Scale Dataset
Jun 21, 2026
MemoryWAM: New World Action Model with Persistent Memory Achieves Breakthrough in Long-Horizon Robot ManipulationAether AI Raises $20M to Build Causal World Models for Embodied Intelligence — A New Paradigm Beyond Scaling LawsGalaxy General Launches AstraBrain-WBC 0.5: The World's First Humanoid 'Cerebellum' GPT Foundation Model