Skip to content
Technical Analysis
Minimal authority may have multiple answers: what must an approval specify?
Research

Minimal authority may have multiple answers: what must an approval specify?

Synthetic enumeration retains two incomparable minimal grants; approval needs the actual footprint and an explicit selection objective.

2026-11-07中文 →

Minimal authority may have multiple answers: what must an approval specify? ​

A task needs one read operation, but each available grant brings a different additional permission. Either grant is sufficient, and neither can be reduced within the available catalogue. Someone must still choose. A perfect minimality result leaves that decision unexplained: which additional capability did the approver accept?

The same-day research includes a small, exhaustive synthetic counterexample. Even after equivalent grants are collapsed, two non-containing minimal footprints remain. The resulting judgment is about mechanism: retain the task, footprint, comparison objective and rationale. It does not estimate how frequently deployed workloads encounter this situation.

Minimal does not mean uniquely least ​

The paper defines a grant's authority footprint as the union of operations it covers and compares grants by set containment. A minimal footprint has no strictly smaller sufficient footprint. A least footprint would be contained in every sufficient alternative. Minimality can therefore leave several incomparable answers. Paper, sections 3, 5 and 7

This also separates names from capabilities. Two names covering the same footprint should be treated as equivalent. Two grants sharing a risk label may still authorize different operations. Merging both situations loses the distinction between aliases for one answer and alternatives carrying different trade-offs.

One read task leaves two answers ​

The internal synthetic experiment uses four grant names: two cover {read, x}, one covers {read, y}, and one covers {read, x, y}. The task requires only {read}. Enumeration yields 15 sufficient grant combinations, collapsing to three distinct footprints. Both {read, x} and {read, y} are minimal. Each contains an operation absent from the other, so containment cannot select between them.

Setting additional-operation costs to x=10, y=1 selects the footprint containing y; swapping those costs reverses the choice. Both selections remain minimal. The difference comes from an explicit objective, rather than a change in sufficiency. Adding a narrow {read} grant as a control produces 31 sufficient combinations and four footprints, with a unique minimum.

These are internal experimental findings, not an independent reproduction. The experiment checks synthetic sets and explicit weights under the paper's definition. It runs no real API requests, model authorization choices or resource conditions. The experiment record retains the inputs, results and limits. It identifies information a score can omit; it cannot establish the correct security preference.

When to ask is separate from what to choose ​

The Research Object also inspected OpenHands' risk-confirmation policy. Its confirmation method returns a boolean, defaults to a high-risk threshold and requires confirmation for unknown risk by default. That answers when to ask. It does not enumerate sufficient grants or choose among incomparable minima. Official confirmation implementation

Our interpretation is that confirmation frequency and selection accountability can both matter. They cannot substitute for each other. Asking every time still leaves an approver uninformed if the prompt shows only a scalar risk label. Conversely, displaying footprints does not determine which requests require human intervention.

An accountable choice record should identify the task, selected actual footprint, feasible incomparable alternatives and the objective—such as narrower reach versus fewer state-changing operations. This is a proposal for preserving decision responsibility, not a tested deployment improvement or a requirement to display every grant combination to a user.

Catalogue proportions are not deployment frequencies ​

The paper reports non-unique minima in 147/493 tasks in one Microsoft permission family and 90/2822 Google resource tasks. Different catalogues and generated task families define those denominators. They do not measure production security incidence across the platforms. The paper's model experiments likewise offer closed choices among sufficient grants; reported minimality rates are not success rates for free-running scope requests. Paper, sections 6, 9, 11 and 14

A counterpoint constrains the story: in the paper's sampled Google workflows, greedily selecting the narrowest available minimum reaches the whole-workflow minimum in 98.8–99.5% of cases. This limits an explanation that attributes all spread to stepwise accumulation. Our counterexample concerns ambiguity at selection time; it does not show that accumulation necessarily causes over-privilege.

Evidence and sources ​

Three evidence roles remain separate. Definitions and numerical results are attributed to the paper. Public code establishes the behavior of the inspected confirmation method. Exhaustive synthetic enumeration is internal evidence. No independent third party reproduced the paper's results here. Citation and publication do not establish validation, and a successful small experiment does not establish general validity.

The main sources are Xin Xu's Incomparable Minima in Agent Tool Authorization, arXiv preprint 2610.31914, version 1, sections 3–14 and OpenHands' official confirmation-policy source. Same-run snapshots are content-hash pinned; the branch snapshot is not a known benchmark-release commit. The internal record includes the inputs, enumeration, weights and control. No author data package was obtained or reproduced.

What changes once resources and conditions enter? ​

Method-level tables omit resource-specific and conditional constraints, and this research measures no deployed task distribution. A finer catalogue can remove some ambiguity, as the synthetic control demonstrates. That does not establish a unique least grant for every real task. Fixed resource-conditional catalogues, actual task footprints and declared risk objectives are needed to determine which alternatives should reach an approver—and who owns the objective.

Last updated: