LLM agent systems often combine multiple security mechanisms like provenance tracking, authorization, and policy enforcement, but these individually correct controls can fail when security context is lost or reinterpreted across component boundaries. This failure, called security-context discontinuity, leads to end-to-end vulnerabilities that local checks miss.
CONTINUITY addresses this by modeling each component with assume-guarantee contracts and carrying authenticated security context across transitions using signed root grants, provenance commitments, and role-bound receipts. The formal end-to-end consequence integrity (ECI) requires every external effect to have a verifiable witness linking the principal, task, provenance, policy, and finality boundary.
In 2,560 parameterized attack instances across 128 fault-domain classes, the full CONTINUITY configuration committed no harmful external effect while completing all benign tasks and escalating ambiguous cases. The results demonstrate that secure agent execution requires explicit contracts preserving guarantees across the entire instruction-to-effect path.