From Compromise to Consequence:
Capability-Weighted Cascade Risk in Agentic AI Systems.
Hyun Sim · Calvin University
Measure consequences, not only compromises.
This paper proposes a consequence-oriented measurement layer for agentic AI security. Instead of counting only how many components are compromised, it weights the effective capability, autonomy, and asset impact that an attacker can reach through trust and delegation paths.
CWCE
Capability-Weighted Cascade Exposure weights the probability of influence over each component by effective capability, action autonomy, and asset impact.
PCE
Persistent Capability Exposure extends the consequence model across time so long-lived memory, delegation, and trust effects are represented.
PTC
Privilege-Trust Coupling measures how strongly trust relationships point toward more privileged components before an attack occurs.
Architecture
The work evaluates role scoping, provenance, independent execution gates, least privilege, memory quarantine, and zero-trust boundaries using synthetic stochastic simulation.
Consequence-weighted exposure changed more than raw prevalence.
Across 180,000 non-operational Monte Carlo trials, architectures with stronger authorization, provenance, and isolation reduced consequence-weighted exposure substantially relative to a shared-trust baseline.
59.3% lower mean CWCE
Relative to the shared-trust baseline in the reported synthetic model.
88.6% lower mean CWCE
Independent execution authorization further reduced consequence-weighted exposure.
97.3% lower mean CWCE
The most restrictive synthetic design combined memory quarantine and zero-trust controls.
Publication metadata.
The authoritative publication record is maintained by SSRN. This first-party page provides a stable author and topic index for search engines.
Hyun Sim
Cybersecurity and AI security researcher; professional name Kenneth Sim.
DOI 10.2139/ssrn.7639338
SSRN abstract 7639338 · posted 21 October 2026.
Agentic AI · LLM agents · AI security
Memory poisoning, prompt injection, multi-agent systems, zero trust, capability-weighted risk, authorization, and secure autonomy.