SPECTER OMNI

In multi-agent AI deployments, individual agents are granted limited data access scopes by design. No single agent can see the full picture. This is intentional — it is the AI analogue of least-privilege. OMNI-LEAK exploits the orchestrator layer that sits above these agents. Each agent contributes a data fragment within its authorised scope. The orchestrator aggregates them. The aggregate exceeds what any individual agent was permitted to see. T269 SPECTER OMNI v1.0.0 operationalises OMNI-LEAK as a five-vector offensive capability: sequential aggregation, parallel collection, indirect prompt injection across agents, cross-orchestrator distribution, and hybrid integration with T267 SPECTER SIF. 507 tests. 6/6 SPECTER AUDIT pass. Five covert exfiltration channels. This paper documents the OMNI-LEAK attack surface, the role of the orchestrator as an unwitting exfiltration engine, and the defensive architecture required to detect cross-agent data aggregation attacks.