AGENTIC-SYSTEMSJuly 14, 20261 min read

A Theory of Least Autonomy in AI

Researchers propose the 'Theory of Least Autonomy' as a necessary evolution of the traditional 'least privilege' principle to secure agentic AI systems from compounding and abusing permissions.

Computer Science > Artificial Intelligence

Title:A Theory of Least Autonomy in AI

View PDF HTML (experimental)Abstract:Least privilege, the principle that an identity should hold only the permissions strictly required for its task, has been a foundational primitive of access control for decades. We argue that this principle is insufficient for agentic AI systems, which do not merely hold permissions but can combine, approve, and amplify them across workflows and system boundaries.

We propose least autonomy as an appropriate generalization and develop a formal theory. First, we define a compositional blast radius d(a,b) that measures structural separation between actions in an enterprise hierarchy, combining an ultrametric tree with lattice-valued confidentiality, integrity, and control-context labels. Second, we define a directed agent influence graph G(theta). An arc from U to V requires a directed shared-resource write-to-read meeting or a conservative undirected agent-to-agent (A2A) communication meeting, and a meeting-conditioned influence potential at or above an externally selected policy threshold theta. A catalogue-radius profile supports calibration and audit of theta. Finally, we define a collusion predicate over graph reachability that detects authorization composition, decision manipulation, and cross-domain capability composition.

Bibliographic and Citation Tools

Code, Data and Media Associated with this Article

Demos

Recommenders and Search Tools

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

Source: arXiv cs.AI Recent

More in this category

agentic-systems

BatteryLake: Agentic, Physics-Grounded Curation of Heterogeneous Battery Aging Data and Benchmarking

Researchers have introduced BatteryLake, a governed data lakehouse that automates the curation of heterogeneous battery aging data using LLM agents and physics-grounded rules. This framework aims to accelerate research in battery health management (SOH and RUL) by providing a standardized benchmark of 41 datasets.

agentic-systems

LegalFarePlan: A Label-Setting Framework for Fare-Transparent Urban Rail Route Planning under Non-Additive Fare Rules

Researchers have proposed LegalFarePlan, a route-planning framework that optimizes urban rail fares by leveraging non-additive fare rules through legal exit-and-reentry operations.

agentic-systems

Task-Conditioned Synthetic Data Generation for Improving Machine Learning Performance in Agricultural Prediction Tasks

Researchers have proposed Task-Conditioned Synthetic Data Generation (TCSDG), a novel algorithm that addresses data scarcity in agricultural machine learning. By combining Bayesian Networks with transformer-based models, TCSDG significantly improves the accuracy of crop yield prediction and crop type classification.

NOW LET US Related – A Dynamic Scene Interaction Reasoning Framework for Scene-level Lane-Change Intention and Trajectory Prediction of Multiple Interacting Vehicles

agentic-systems

A Dynamic Scene Interaction Reasoning Framework for Scene-level Lane-Change Intention and Trajectory Prediction of Multiple Interacting Vehicles

Researchers have proposed DSiGAT, a novel dynamic scene graph attention framework that predicts both the lane-change intentions and future trajectories of multiple interacting vehicles simultaneously, significantly improving autonomous driving safety.

agentic-systems

Scaffolding the Strategist: Architecture-Dependent Reasoning Interventions in Hotelling Spatial Markets

A new study reveals a significant crossover interaction between reasoning scaffolding and LLM architecture, showing that interventions helping standard models can degrade reasoning-optimized models.

agentic-systems

Coresets Before Score Sets: Evaluation-Unsupervised Prompt Subset Selection for LLM Benchmarks

Researchers propose an evaluation-unsupervised prompt subset selection method to compress massive LLM benchmarks while preserving evaluation accuracy, significantly reducing computation costs.