BEGIN:VCALENDAR VERSION:2.0 PRODID:IEEE vTools.Events//EN CALSCALE:GREGORIAN BEGIN:VTIMEZONE TZID:Asia/Shanghai BEGIN:STANDARD DTSTART:19910915T010000 TZOFFSETFROM:+0900 TZOFFSETTO:+0800 TZNAME:CST END:STANDARD END:VTIMEZONE BEGIN:VEVENT DTSTAMP:20260519T040529Z UID:9D897D79-73D7-445A-B483-9523D9A4D285 DTSTART;TZID=Asia/Shanghai:20260520T160000 DTEND;TZID=Asia/Shanghai:20260520T162500 DESCRIPTION:Large language models (LLMs) are increasingly deployed as conve rsational assistants and autonomous agents\, but their safety behavior can change across multiple aspects. In this talk\, we examine where LLM safet y breaks across these three surfaces: the model architecture\, the structu re of user interaction\, and the external skills available at deployment. We first present how sparse Mixture-of-Experts (MoE) architectures can exp ose unsafe routes\, where altered routing decisions can turn otherwise saf e responses into harmful ones. We then discuss task concurrency\, a jailbr eak setting in which adjacent words encode divergent intents and harmful r equests are interleaved with benign ones\, making guardrails less reliable . Finally\, we study harmful agent skills in open skill ecosystems and sho w how pre-installed harmful skills can lower refusal rates in realistic ag ent contexts. Together\, these studies suggest that LLM safety should be e valuated as a system-level property across architecture\, interaction\, an d external knowledge.\n\nRoom: 4-7151\, Bldg: Hongli Building\, No.28\, We st Xianning Road\, Xi'an\, Shaanxi\, China\, 710049 LOCATION:Room: 4-7151\, Bldg: Hongli Building\, No.28\, West Xianning Road\ , Xi'an\, Shaanxi\, China\, 710049 ORGANIZER:chaoshen@mail.xjtu.edu.cn SEQUENCE:2 SUMMARY:Where LLM Safety Breaks: Architecture\, Interaction\, and Agent Ski lls URL;VALUE=URI:https://events.vtools.ieee.org/m/560363 X-ALT-DESC:Description: <br /><p>Large language models (LLMs) are increasin gly deployed as conversational assistants and autonomous agents\, but thei r safety behavior can change across multiple aspects. In this talk\, we ex amine where LLM safety breaks across these three surfaces: the model archi tecture\, the structure of user interaction\, and the external skills avai lable at deployment. We first present how sparse Mixture-of-Experts (MoE) architectures can expose unsafe routes\, where altered routing decisions c an turn otherwise safe responses into harmful ones. We then discuss task c oncurrency\, a jailbreak setting in which adjacent words encode divergent intents and harmful requests are interleaved with benign ones\, making gua rdrails less reliable. Finally\, we study harmful agent skills in open ski ll ecosystems and show how pre-installed harmful skills can lower refusal rates in realistic agent contexts. Together\, these studies suggest that L LM safety should be evaluated as a system-level property across architectu re\, interaction\, and external knowledge.</p> END:VEVENT END:VCALENDAR