CVE-2026-59891 Regression & Control Lab

An isolated hands-on lab and read-only audit project that compares the actual vulnerable and fixed versions of @sigstore/oci with the same input and checks environment-specific exposure conditions.

The CVE Reporter for this project is gyubin02.

Recruitment/interview summary: CVE-2026-59891 vulnerability management automation case

What this verifies

The core of CVE-2026-59891 is hostname confusion that occurs when selecting Registry credentials.

0.7.0: Because cr.io is contained within the ghcr.io string, it may select ghcr.io credentials.
0.7.1: After normalizing the hostname, it selects only exactly matching credentials.

This repository first automatically verifies two credential selection scenarios.

collision: Only ghcr.io credentials exist, but the target is cr.io
exact: There is a cr.io credential that exactly matches the target

The vulnerable version must incorrectly select credentials in the first scenario, and the fixed version must reject them. Both versions must succeed in the second, normal scenario.

We also added a dynamic comparison that uses a one-off mock Registry at 127.0.0.1 instead of a real external Registry. The vulnerable version 0.7.0 makes 2 local requests, and after the authentication challenge a synthetic Authorization header is observed on the second request. The fixed version 0.7.1 is rejected at the credential selection stage, resulting in 0 requests. Only the observation flag and request count remain in the results, instead of header values.

Project exposure audit CLI

cve-2026-59891-audit reads only the metadata of the target project's package-lock.json and an explicitly specified Docker config to distinguish the following:

Affected @sigstore/oci installation locations and versions
Whether a vulnerable version exists and whether real exposure conditions are met
Registry substring collisions and the actual selection target according to JSON key order
Environment-specific priorities, remediation actions, and retesting procedures
JSON and Markdown evidence with Registry hostnames and absolute paths removed

Run directly with the synthetic fixture:

npm run audit:demo

Audit another Node.js project:

npm run audit:project -- \
  --project ../target-project \
  --docker-config ../review-copy/config.json \
  --image cr.io/example/demo \
  --destination-trust unknown \
  --format markdown \
  --output reports/cve-2026-59891.md

--destination-trust must explicitly specify one of the following:

trusted: Confirmed that destination is restricted by code and an allowlist
untrusted: External input or workflow input affects destination
unknown: Not yet confirmed

Verdicts are divided into exposure_conditions_met, potential_exposure, affected_component_only, not_detected, and indeterminate. not_detected is only a result within the recorded input scope, not a statement of "safe" or "compliance".

Safeguards

The regression lab does not read or modify the actual ~/.docker/config.json.
Uses only an isolated, dedicated HOME and worthless lab-user / LAB_ONLY_FAKE_TOKEN.
Calls the package only when HOME, the lab marker, and the fake values are all correct.
The existing selection regression probes block all HTTP(S), TCP, TLS, DNS, and fetch calls as tripwires.
On the Node network surface used by the dynamic probe, only HTTP requests to the exact OS-assigned 127.0.0.1 port are allowed. External and different-port HTTP, DNS resolver, TLS, UDP, global fetch, and WebSocket are blocked.
Does not inherit the current session's token or Docker environment variables into child processes.
Tests that stdout/stderr and audit results store nothing, not even fake headers, and keep only booleans and request counts.
The audit CLI cannot omit --docker-config and does not infer files from HOME or DOCKER_CONFIG.

Because it intentionally installs a vulnerable package, npm package publishing is blocked with "private": true. Do not use it as a dependency of production code. For this reason, npm audit reporting 0.7.0 is an expected result, and the scope and reason for the exception are documented in SECURITY.md.

Execution

Requirements: Node.js 22.22.2+, 24.15.0+, or 26 or higher. The recommended version is pinned in .nvmrc.

npm ci --ignore-scripts
npm test
npm run demo
npm run dynamic:demo
npm run audit:fixture
npm run audit:demo
npm run evidence

Expected key output of npm run demo:

0.7.0  collision  credential-selected
0.7.1  collision  credential-rejected
0.7.0  exact      credential-selected
0.7.1  exact      credential-selected

Regression result: PASS

Expected key output of npm run dynamic:demo:

{
  "vulnerable": {
    "packageVersion": "0.7.0",
    "credentialSelection": "credential-selected",
    "requestCount": 2,
    "authorizationObserved": true,
    "digestVerified": true,
    "networkPolicy": "exact-loopback-only"
  },
  "fixed": {
    "packageVersion": "0.7.1",
    "credentialSelection": "credential-rejected",
    "requestCount": 0,
    "authorizationObserved": false,
    "digestVerified": false,
    "networkPolicy": "exact-loopback-only"
  },
  "regressionResult": "PASS"
}

There are 34 automated tests in total. npm run evidence generates a total of 7 files: 6 audited artifacts plus SHA256SUMS to verify their integrity.

Capabilities this project demonstrates

Describes at the code level the conditions under which a real CVE occurs and safely reproduces it
Regression test design comparing vulnerable and fixed versions under identical conditions
Loopback-only test design that dynamically compares even whether synthetic credentials are transmitted locally
Assesses real-world attack prerequisites separately from the public CVSS
Statically identifies package-lock v1, v2, v3, aliases, and nested dependencies
Automatically generates JSON and Markdown audit evidence excluding secrets
Vulnerability management that flows from detection → impact analysis → fix verification → controls
Fail-safe security testing that never touches real credentials

Sources and scope

This project is a limited educational controlled assessment. The loopback results only demonstrate behavioral differences in a synthetic environment; they do not prove credential exposure or compromise, or control effectiveness, in real enterprise environments. It does not imply CISA certification, ISMS-P certification, or compliance with any specific organization. The Node API guard is a defense layer for observed execution paths of locked dependencies, not an OS-level network namespace or firewall.

Detailed scope, tests, risk, and control documentation is in docs/, and npm run evidence generates reproducible execution evidence and a SHA-256 manifest.

Beliebteste

cve-2026-59891-control-lab