一作ACL 2026Findings· Findings of ACL
SLIP : Soft Label Mechanism and Key-Extraction-Guided C o T -based Defense Against Instruction Backdoor in API s
Wu Zhengxian · Juan Wen · Wanli Peng · Haowei Chang · Yinghan Zhou · Xue Yiming
摘要
Customized Large Language Model (LLM) agents face a critical security threat from black-box instruction backdoors, where malicious behaviors are covertly injected through hidden system instructions. Although existing prompt-based defenses can often detect poisoned inputs, they generally fail to r...