Google dropped a paper on the ai attack surface in 2026. If you’re building multi-agent systems these should be embossed into your frontal lobe…
Congestion Traps: multi agent DDOS
Interdependence Cascades: compounding hallucination across interdependent agents.
Tacit Collusion:
Using the env to gaslight multiple agents into conforming
Compositional Fragment Traps:
Fragmented payload that triggers on multi-agent aggregation
Sybil Attacks:
Fabricates multiple pseudonymous agent identities to disproportionately influence decision making