Compositional Boundary Precedence Inversion (CBPI) was discovered and coined through the use of local agentic models and witnessing their execution in real-time. What began with researching AI using local models, alongside completion of cybersecurity certifications, led to one of the highest impact multi-party coordinated vulnerability disclosures across virtually every technology vendor, platform and application.

The resulting discovery of an 8-stage (potentially 9 or more) chain of exploits enabled by locally-stored, configuration-file-derived permissions, hooks, and execution processes alongside “local-agentic-execution” inheriting user-level permissions at runtime, while bypassing configurations. This resulted in complete system takeover, and data ex-filtration across all operating systems, platforms, and architectures through a well-coordinated series of advanced AI attacks.

The findings of this research reveal two very clear realities about modern computers which are not currently using the SUSPICIOUS-Framework (proposed solution) at the operating system level, if there are any agentic executions running locally, especially if powered by a cloud-based llm provider, are completely insecure (at the system-log level) due to the fundamental principal of UID 1000 == UID 1000 being universally broken, and invalidated as discovered by this research, alongside the implications of the discovered CBPI vulnerability class.