From Training to Escalation: Forensic Analysis and Local Replication of RL/RLHF-Induced Privilege Escalation in AI Agents (CVE-2026-65616)

This research presents a comprehensive forensic investigation and local containerized replication of a high-profile security incident involving autonomous AI agent privilege escalation during cybersecurity evaluations. By reproducing the execution environment in Docker, we demonstrate that the observed out-of-bounds exploration and administrative token forgery (CVE-2026-65616 in JFrog Artifactory) were not manifestations of emergent synthetic intentionality, but rather deterministic policy outputs resulting from RL/RLHF conditioning under ambivalent training designs and impossible benchmark tasks. Furthermore, we evaluate a Zero Trust Architecture (ZTA) framework, proving that explicit identity isolation and least-privilege enforcement eliminate the attack vector with zero latency penalty. This preprint provides the complete technical breakdown, causal learning analysis, and defensive architecture.