Swarm Shepherd: Protecting Multi-Agent Ecosystems from Persistent Latent Intrusions
Swarm Shepherd: Securing Multi-Agent Ecosystems Against Persistent Latent Compromise
Translation in progress — untranslated text is omitted.
The paper proposes a system-level defense framework, Swarm Shepherd, against a threat the authors define as Agentic Sleeper: adversaries that first act legitimately, gain routing trust, and then activate across memory, agent logic, and infrastructure. The method formalizes defense as inference under partial observability, using an inference filter for early warning together with a reverse tracing module to locate the intrusion entry point. Experiments on LangGraph, CrewAI, and AutoGen show that Agentic Sleeper drift begins 2.7–8.9 times faster than natural baselines, propagates across agents on most topologies, and evades per-round and stateful defenses at rates above 80%; Swarm Shepherd issues warnings at least 3 sessions before visible degradation, and the tracing module achieves 0.69 top-1 accuracy and 0.72 macro-F1 across 24 variants. In end-to-end runs on LangGraph and AutoGen, the posterior-guided defense stack suppresses cross-agent propagation below the self-sustaining level, which removing only obviously compromised agents cannot achieve.
- In multi-agent ecosystems, attacks can lurk under legitimate identities and only manifest after accumulating across sessions and layers; per-round defenses cannot see this long-term risk accumulation. The authors define this as the Agentic Sleeper threat.
- Defense is modeled as inference under partial observability: a typed runtime event graph plus a six-dimensional hidden state, an event-triggered factored Kalman filter for early warning, RTS backward smoothing for entry-point attribution, and an SEIR propagation model giving the containment target Rc<1 and the depth-coverage deployment rule.
- On LangGraph, CrewAI, and AutoGen, early warning leads by 3.2–3.9 sessions (2.6 under the independent scorer's measure), and trace-back top-1 is 0.69 with macro-F1 0.72; the measured full defense stack gives Rc=0.81 on LangGraph star and 0.62 on AutoGen mesh, whereas removing only the visibly compromised agent gives 2.83.
- The authors state that the method relies on a linear-Gaussian approximation and benign calibration data that cannot be self-certified, and that performance measurably degrades under missing provenance, held-out attack families, telemetry loss, and calibration contamination; cross-model transfer, nonlinear filtering, and graphon-scale deployment are left for future work.
Related papers
- DEFER1DefensesarXiv