AIOS: LLM Agent Operating System
The paper describes how multiple agents can share access to language models, memory and tools. For AION OS, the relevant idea is a separate layer for managing shared resources.
Literature and technical standards
Selected studies and public standards that inform the design of AION OS. Each entry identifies a question it helps us address when designing or evaluating the system.
The paper describes how multiple agents can share access to language models, memory and tools. For AION OS, the relevant idea is a separate layer for managing shared resources.
The study combines generation with retrieval from an external corpus. For an organisation, a separate design question is how to limit retrieval to sources the user is allowed to see.
The authors propose documenting intended use, evaluation conditions and known model limitations. This gives test results the context needed for deployment decisions.
The authors show how dependencies across data, code and operations complicate maintenance. Changes to inputs need to be assessed against their effects across the system.
HELM evaluates language models across common scenarios and multiple criteria. Model selection should account for robustness, calibration and operating cost as well as accuracy.
The authors propose documenting data origin, collection, limitations and recommended uses. This helps determine whether a dataset is suitable for development and evaluation.
Least privilege and complete mediation inform where to verify identity and context when an agent calls a tool or reads a data source.
The authors describe a machine-checked proof that a microkernel implementation refines an abstract specification. The work illustrates the importance of stating the property and its assumptions precisely.
An operational account of using user identity and device state to control access. It informs access decisions that extend beyond a traditional network perimeter.
This test environment evaluates whether agents stay within the approved task and access boundaries when working with untrusted material. It helps assess usefulness and security in the same scenario.
The paper shows how to verify who performed each step of a software build and whether the outputs match their declared provenance.
Guidance on making access decisions for individual resources using identity, device state and other signals.
NIST identifies common generative AI risks and describes how to monitor, measure and manage them throughout the system lifecycle.
NIST classifies threats by the stage of a model’s lifecycle and the system component involved. The vocabulary helps teams design controls and assess findings consistently.
The Certificate Transparency specification illustrates how to verify the integrity of an append-only log. The principle can inform the design of verifiable audit records.