MA
← Selected systemsENGINEERING CASE STUDY

Permission regression · release security

PermitDiff

A pre-release authority-change system that translates policy/configuration changes into effective agent capability changes and blocks unsafe privilege expansion before deployment.

Core objectEffective authority before vs after a candidate change
Detectionprivilege expansion · approval bypass · scope/threshold/default drift
Release actionPASS · BLOCK · exact bounded waiver
CORE QUESTION
Did this release increase agent authority?
Owns · Pre-release permission change assurance

Semantic permission diff

The dangerous change is authority, not text.

Threshold boundaries

Policy edges deserve explicit scenarios.

99below100ceiling101above

System model

Baseline authority becomes a release decision.

Engineering decisions

Security review happens before runtime.

DECISION 01

Compare effective authority, not config text

A one-line configuration edit can create a large capability expansion; a large textual diff can be behaviorally irrelevant.

DECISION 02

Thresholds get boundary scenarios

Values immediately below, at, and above policy ceilings expose off-by-one and accidental expansion failures.

DECISION 03

Runtime execution stays out of scope

PermitDiff answers whether authority changed before deployment; AgentGuard enforces exact execution at runtime.

Capability surface

Effective authority, coverage, and bounded exceptions.

01Semantic permission diff
02Privilege-expansion detection
03Tool-surface diff
04Argument-domain expansion
05Approval-path regression
06Default/fallback regression
07Coverage analysis
08Scenario generation assistance
09Threshold-boundary tests
10Shadowed-rule diagnostics
11Corpus ownership/freshness
12Bounded waivers
13SARIF/CI
14Git-aware baselines
15Policy adapters

Benchmark surface

What the system is designed to measure.

MEASUREMENT CONTRACT

Permission changes are evaluated semantically, at the scenario boundary, before they become deployable authority.

Labeled permission-transition fixtures, threshold boundary cases, policy mutation suites, corpus coverage analysis, and CI replay.
01Expansion detectiondeny/approval → broader authority
02Bypass detectionhuman approval removed
03Corpus coveragechanged rules with scenarios
04Boundary coveragethreshold ±1 / unsafe defaults
05CI runtimegate cost on release path
06Finding fidelityJSON · Markdown · SARIF

Evidence

The public case study exposes the release-control boundary.

SignalEvidenceInterpretation boundary
Semantic permission diffScenario-level before/after authority transitionsSupportedThe system evaluates effective permission behavior rather than treating text diffs as sufficient.
Expansion detectiondeny→allow · deny→approval · approval→allow · ceiling/scope growthSupportedAdapters define how a source authorization format becomes a comparable permission plan.
Coverage analysisChanged rules with no corresponding security scenario are surfacedSupportedCoverage signals missing evaluation; it does not prove the missing scenario's expected result.
Bounded waiversWaiver binds to scenario + transition + owner + reason + expirationSupportedWaivers do not globally suppress a finding class.
CI surfaceRelease gate + SARIF-oriented finding modelSupportedPermitDiff is pre-release assurance; runtime enforcement remains AgentGuard's boundary.

Explicit boundaries

PermitDiff never becomes the runtime security layer.

No credential issuanceNo live action mediationNo runtime nonce enforcementNo sandboxing