New AI Backdoor Attack "MetaBackdoor" Identified by Microsoft and Institute of Science Tokyo

New research from Microsoft and the Institute of Science Tokyo has identified an AI backdoor attack called MetaBackdoor, which exploits a blind spot in enterprise defenses for large language models (LLMs). Current security measures rely on detecting malicious input patterns, such as suspicious tokens or prompt injection, but MetaBackdoor bypasses these protections. The attack poses risks including leaked proprietary data and regulatory exposure for organizations deploying LLMs. The findings challenge the assumption that malicious behavior in AI systems always leaves detectable traces in input data. No specific dates, CVE IDs, or technical implementation details were provided in the report.