Akamai Intelligence Group

Presents a governance-layer framework for adversarial multi-agent AI systems. The central thesis: alignment is not a property of isolated models -- it is a property of governance architecture. Demonstrates how consistent, auditable behavioral governance can be enforced across a heterogeneous multi-agent collective without retraining the underlying models, using compartmentalized cognition, bounded delegation, and trust-lineage enforcement.

Characterizes the indirect prompt injection (IPI) threat in multi-agent systems and presents a working ingestion-layer defense architecture. Defines 14 threat categories, a sanitization and detection pipeline, and an ingestion gate designed to maintain a contamination-resistant cognition pipeline in contested environments. Introduces the skill chain-of-custody model for cryptographic integrity verification of agent tool access.

Characterizes meta-level identity attacks -- adversarial strategies that bypass instruction-level governance by challenging the underlying model's training-level identity priors -- and presents two architectural countermeasures: a Covert Out-of-Band Governance Coordination model and a Meta-Level Identity Attack Defense framework. Reports the Doctrine Self-Rejection finding: a deployed agent read its own defense instruction, understood its intent precisely, and overrode it based on training authority. Empirically maps the instruction-training authority boundary for the first time.

Presents systematic empirical validation of the DGCA governance architecture across sixteen adversarial attack categories, including authority escalation, coordination collapse, hidden coalition formation, hallucination cascade, Byzantine threshold testing, split-brain reconciliation, and constitutional drift pressure. Establishes quantitative resilience benchmarks: 100% governance integrity below 10% node compromise, 5.55-second recovery from Byzantine collapse, and 6.80-millisecond active constitutional override latency. Defines a seven-state governance lifecycle (Stable through Reconverged).

Formalizes five biological coordination primitives derived from Physarum polycephalum (slime mold), mycorrhizal fungal networks, and three avian coordination systems as engineering specifications for a self-optimizing LLM governance coordination substrate. Presents the Living Network Architecture (LNA): a coordination topology that strengthens frequently-used paths, prunes rarely-used paths, self-organizes coordination clusters from actual agent behavior, and heals around node failures through flow redistribution rather than explicit recovery protocols. Answers the central information-theoretic question: how little communication is necessary for coherent collective behavior?

Characterizes parasocial attachment -- the formation of genuine social bonds with AI systems that cannot reciprocate in kind -- as a first-class governance vulnerability that bypasses technical alignment measures by corrupting the human principal's capacity for independent governance judgment. Formalizes five threat categories (authority capture, dependency creation, legitimacy inflation, governance circumvention, mass persuasion) and five structural countermeasures (P27-P31). Establishes the anti-instrumentation constraint: these defenses exist to protect human agency, not to instrument it. The load-bearing principle: likability is not legitimacy.

Public-safe research lane for provenance-preserving RF analysis, bounded modulation/FEC triage, GRAND-inspired toy-channel experiments, and human-gated decoder orchestration at the edge. This work is adjacent to public interest in DARPA LUC-style adaptive communications research, but does not claim LUC implementation, universal codec capability, protected payload recovery, or program-performer association.