Work

AISE: Enforcing Explicit Assembly and Release Authorization Boundaries in Multi-Agent LLM Systems

Year: 2026
Type: preprint
Abstract: Modern Large Language Model (LLM) agents increasingly operate as multi-agent pipelines that retrieve, summarize, and combine data across untrusted external sources and sensitive internal systems. While existing security architectures—such as tool-call gateways and interpreter-based Information Flow Control (IFC) frameworks like CaMeL—restrict which tools an agent may invoke or how structured variables propagate, they face a critical blind spot during unstructured multi-agent synthesis. This paper introduces AISE (Architecture for Isolated, Scoped Execution), a capability-based security model built around a single invariant: permission to access does not imply permission to combine, and permission to combine does not imply permission to release. AISE enforces this invariant through a cryptographic Token Spine of short-lived, holder-bound, monotonically attenuated capability tokens, paired with two deterministic policy checkpoints: an Authorize Assembly Gate and an Authorize Release Gate. We formalize the AISE pipeline, contrast it with prior art, and present empirical results from an 8-case adversarial stress test demonstrating 100% deterministic containment against privilege escalation, transitive taint laundering, token replay, and steganographic declassification smuggling. Code (aise_prototype.py) is licensed under Apache 2.0; see LICENSE.
Language: English
Open Access status: green
OpenAlex ID: W7235011856
Cites:
Cited by:
Related to: 10