Evidence-linked developments
- Distributed backdoors in multi-agent LLM systems bypass local monitoringFactual summary
New research identifies 'distributed backdoors' in multi-agent LLM systems, where harmful payloads are split across agents, allowing them to bypass local runtime monitors and create compositional harm.