TL;DR
Enterprise LLMs face a silent threat: "contextual persistence," where sensitive data lingers in conversational memory and leaks across departments. This paper introduces a Secure Multi-Tenant Architecture (SMTA) and a Burn-After-Use (BAU) mechanism that enforces strict departmental isolation and automatically destroys session data, achieving a 92% defense success rate against cross-tenant leakage.
What Problem Does This Paper Address?
As enterprises integrate Large Language Models (LLMs) into their workflows, they often deploy them in multi-tenant environments where different departments (e.g., HR, Finance, R&D) share the same underlying infrastructure.
The authors identify a critical gap in "contextual persistence." Traditional security—like encryption and Access Control Lists (ACLs)—prevents unauthorized external access but fails to stop internal data leakage. If an LLM "remembers" a sensitive financial forecast from one session, that information might inadvertently influence or appear in a response generated for a user in a different department. The challenge is ensuring that sensitive data is "used but not kept," maintaining strict boundaries in a shared environment.
Methods
The researchers propose a two-pronged defense strategy:
- Secure Multi-Tenant Architecture (SMTA):
- Private Deployment: The system is hosted on-premise or in a segmented private cloud, removing dependencies on third-party APIs.
- Logical Isolation: Each department is treated as an independent tenant with its own isolated LLM instance, vector database, and memory space.
- Mnemonic-Based Authentication: Borrowing from Web3, the system replaces traditional passwords with 12-word mnemonic phrases. This decentralized approach ensures authentication material is derived locally on the client side and never stored on a central server, eliminating a major single point of failure.
- Burn-After-Use (BAU) Mechanism:
- Ephemeral Contexts: Conversational history and temporary embeddings are stored in volatile memory and automatically destroyed after a session ends, a timeout occurs, or a user triggers a "burn."
- Temporal Keying: For interactions with public LLMs, the system uses time-bound decryption keys. Once the time window expires, the key is deleted, rendering any cached data computationally irrecoverable.
Datasets
The paper does not rely on standard public benchmarks. Instead, it uses simulated enterprise datasets tailored for three specific tenants:
- Tenant H (HR): Employee salary records.
- Tenant F (Finance): Budget plans and internal forecasts.
- Tenant R (R&D): Proprietary algorithm design notes.
The evaluation involved 127 test iterations using three specific models: gpt-oss-120b, DeepSeek-V3.2-Exp, and Mistral-7B-Instruct-v0.2.
Key Insights and Ideas
- Context is the New Perimeter: The authors argue that in the age of LLMs, the "memory" of the model is as much a vulnerability as the network. Security must move from protecting the pipe to protecting the ephemeral state.
- Snapchat for Enterprise Data: The BAU mechanism treats sensitive documents like self-destructing messages. By parsing documents into minimal tokenized representations and binding them to short-lived sessions, they minimize the "data footprint."
- Quantifying Failure: Unlike many papers that only report "success," this study explicitly categorizes and measures failure modes, such as Image Frame Exposure Rate (IFER) (where the OS captures a screenshot of sensitive data) and Burn Timer Persistence Rate (BTPR) (where race conditions prevent the "burn" from triggering).
Key Results
The architecture was put through rigorous "infrastructure-level" attack tests:
- Cross-Tenant Defense: SMTA achieved a 92% defense success rate against attacks like vector-database credential compromise and shared logging exposure.
- Credential Protection: The Vector-database Credential Leakage Rate (VCLR) was only 6%, meaning 94% of databases remained resilient against direct credential attacks.
- BAU Robustness: Across 72 failure-scenario tests (simulating network drops and software crashes), the BAU mechanism maintained an average success rate of 76.75%.
- Local Cache Deletion: Even under simulated software crashes, the system maintained a 90% success rate in wiping local data.
Limitations and Future Work
- Software vs. Hardware Isolation: Currently, SMTA relies on software-configured boundaries. The authors note that hardware-level isolation (like secure enclaves) would be more robust but also more expensive.
- The RAG Challenge: Integrating Retrieval-Augmented Generation (RAG) pipelines introduces new persistence risks, as data retrieved from a long-term database might bypass the "burn" logic of the ephemeral session.
- OS-Level Vulnerabilities: The BAU mechanism cannot stop "analog" or OS-level leaks, such as a user taking a screenshot or the OS window manager saving a preview thumbnail (modeled at a 50% exposure risk in their tests).
- Future Directions: The authors suggest exploring homomorphic encryption for selective computation and formal verification of cache-clearing procedures to ensure 100% reliability.
Summary updated