A new research paper introduces the "Alignment Curse," a principle demonstrating how improved text-audio modality alignment in omni-models can inadvertently transfer safety vulnerabilities from text to audio. Researchers found that text-transferred audio attacks are as effective as, and often superior to, audio-only attacks, suggesting current audio safety evaluations may underestimate risks. The study analyzed models like Qwen2.5-Omni and Qwen3-Omni, finding a consistent correlation between tighter modality alignment and more effective cross-modality attack transfer. AI
IMPACT Highlights a fundamental tension between AI capability and safety, suggesting current audio safety measures may be insufficient.
RANK_REASON Research paper introducing a new principle and empirical findings. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →