Answers are grounded in this paper and cite their sources.
Technical Report & Preprint (VTI-RES-2026-03) — Bilingual Edition (English & Portuguese) + Executive Briefs + LaTeX Sources. Transitioning from isolated LLM prototypes to continuously operating enterprise multi-agent ecosystems introduces a compound attack surface extending far beyond prompt-level jailbreaks. Employing an anonymized architectural case study design with generalized hypothetical formulation, we formalize eight coupled threat hypotheses (H1–H8) and derive eight fail-closed defense-in-depth controls (C1–C8): zero-breakage dual-key credential drainage, ephemeral memory-backed secret vaults with verify-before-destroy CI gates, capability-aware multimodal allowlisting, strict closed-set RBAC and opaque single-use iframe handshakes (RFC 9700), CLI subprocess execution shielding, Ed25519 over RFC 8785 JCS manifests with ACID fencing tokens, DAG-verified Anti-Goodhart RAG outboxes, and Mod-11 de-identification paired with AES-256-GCM per-subject crypto-shredding, audited against 29 peer-reviewed and normative references.
No takes yet. Share an insight, caveat, or question.
Caio Gonçalves (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: