Case · DiggingBeagle record

Minimum agent API permissions can be incomparable rather than rankable

A study of Microsoft Graph and Google API authorization structures finds many tasks with several incomparable minimal permission sets. Frontier models often choose sufficient but dominated grants in the tested cases.

Scope

The percentages refer to tasks generated from API tables and sampled decisions; they are not rates of actual excessive permissions in production customers or known exploited OAuth grants.

Assessment profile

Assessment method
Setting
evaluation
Exploitation
demonstrated
Evidence
Grade B
Remediation
not applicable
AI role
AGAINST AI
Basis and provenance

Enrich update: CLM-ENRICH-20261027-02 cites SRC-2026-CP21-APIAUTH-INCOMPARABLE-GRANTS at Abstract, model decisions, four risk measures and dominated grants. New evidence concerns the bounded reported mechanism, prerequisite or outcome; setting evaluation; exploit status demonstrated; remediation status not_applicable. No independently comparable impact scale, affected production population or universal exploit success measure was established, so the impact band remains null. Original Claims remain individually attributable and any observed incident or simulated result is distinguished in the narrative.

Assessed Oct 27, 2026 using diggingbeagle.assessment/1.

At a glance

Timeline

  1. Oct 22, 2026
    disclosure

    Public disclosure

Claims & evidence

CLM-ENRICH-20261027-02In 1,708 modeled grant choices, the six evaluated frontier models picked a minimally sufficient option 99.0 percent of the time, but 22.1 percent of choices were dominated on every one of the authors four reported risk measures.supported

Basis: reported finding

Link to claim
CLM-CP21-APIAUTH-INCOMPARABLE-GRANTS-1The paper examines 22,305 Microsoft Graph bindings and 23,852 Google API methods, reporting incomparable minimal grants on 29.8-33.3 percent of generated Graph tasks and 3.2 percent of Google tasks.supported

Basis: reported finding

Link to claim

Implications

The case challenges one-number permission rankings, not the principle of least privilege itself.

Controls and mitigations

  • Use a partial-order comparison of permissions and evaluate alternative minimal grants against task-specific consequence dimensions before user consent.

Unknowns and contradictions

  • The percentages refer to tasks generated from API tables and sampled decisions; they are not rates of actual excessive permissions in production customers or known exploited OAuth grants.

Sources and citation

Material revision history

  1. Oct 27, 2026 · Canonical change recorded · new in release · revision 196