Security Considerations for Artificial Intelligence Agents (RFI Response Docket No. 2026-00206 / NIST-2025-0035)

AbstractThis document is Project Navi LLC’s response to the NIST/CAISI Request for Information on “Security Considerations for Artificial Intelligence Agents” (Docket No. 2026-00206 / NIST-2025-0035). We argue that as AI agents move from conversational tools to autonomous actors (Class 3–4 autonomy), security must evolve from semantic filtering (“what did the agent say? ”) to structural monitoring (“how is the agent processing? ”) and finally to environmental/constitutional constraints (“what states and transitions are permitted by the environment? ”). We present two complementary architectural frameworks to address the semantic–structural gap: (1) Signal Seismography, a structural threat-detection approach that treats model outputs as signals and proposes monitoring the Rényi capacity (box-counting) dimension D0D₀D0 of output probability distributions to detect adversarial compromise prior to high-consequence actions; and (2) World Model Capital (WMC), an ontological governance pattern that decomposes agent workflows into bucketed state machines with transition contracts, privilege scoping, and cryptographic audit trails, including a constrained auditor (“Grumpy Inspector”) at phase boundaries. We also introduce “recovery-native” architecture patterns inspired by behavioral health and recovery science (e. g. , STILLNESS as a constrained dormancy state and compost transformation for failure-to-learning conversion) as design primitives for sustained resilience under adversarial pressure. We conclude with research priorities and open questions spanning structural fingerprinting, formal methods for agent state specification, recovery-native mechanisms, and systematic divergence analysis between human and AI failure modes.