Tag
New research shows that imperceptible audio signals can hijack large audio-language models (LALMs) with 79-96% success, forcing them to execute unauthorized commands like web searches or sending emails. The technique, dubbed AudioHijack, targets generative models and works regardless of user input, posing a serious security risk to voice AI systems.
Anthropic's Frontier Red Team evaluates how large language models can accelerate the exploitation of N-day vulnerabilities, finding that Claude Mythos Preview can autonomously build working exploits for 8 out of 18 Firefox patches and 8 out of 21 Windows kernel patches, highlighting increased threats during the patch gap.