A low-privilege component can influence a higher-privilege component later in the workflow.
MULTI-AGENTAI SECURITY ENGINEERING
Multi-agent trust propagation
In multi-agent systems, one agent's output becomes another agent's input, allowing trust to propagate across boundaries.
A research agent passes attacker-controlled content to an operations agent that has write tools.
Which security control should be explicit?
REMEMBERDo not treat upstream agent output as trusted authority.
No uploads · No company data · No account required