ZeroDark
2026 / External pentestIllustrative sample

External penetration test

Illustrative external penetration test and retest report for Cluckin’ Bell, the fast-food chain located in Paleto Bay. The engagement, assets, observations and historical results are fictional. Threat intelligence references describe real groups.

Executive summary

The internet-facing VPN accepted a password as the only authentication factor. A credential lost to phishing could therefore put remote network access at risk. Testing also found reflected cross-site scripting on the support portal and a live API credential in public GitHub history. Prioritize phishing-resistant MFA and key revocation, then fix the portal and remaining public exposures.

Five issues were confirmed in the initial assessment (two High, two Medium and one Low). The November 11 retest verified three fixes, including both High findings. The support portal XSS and VPN TLS configuration remain open.

Scope

The client supplied four public IPv4 addresses and the domain chuckin-bell.com. Authorization covered the four approved hosts listed below and read-only review of the client’s designated public GitHub repository. The VPN assessment used one approved test account.

  • 192.0.2.10 · vpn.chuckin-bell.com · Remote access gateway
  • 192.0.2.20 · support.chuckin-bell.com · Customer support portal
  • 198.51.100.30 · staging.chuckin-bell.com · Staging web service
  • 203.0.113.40 · www.chuckin-bell.com · Public retail site

Exclusions included employee phishing, denial of service, internal network access, gift card transactions, and use of any discovered credential. The public GitHub review covered repository content and history. Anthropic infrastructure was outside scope.

Methodology

  1. Scope reconciliation and reconnaissance. The tester compared the client’s asset list with public DNS and certificate records. Publicly visible service and repository evidence was recorded against the approved inventory. Any additional assets required client approval before assessment.
  2. Threat-informed review. The tester selected three perspectives based on Cluckin’ Bell’s public exposure, gift card business and dependence on remote access. The team used public threat intelligence to identify relevant control questions for each perspective.
  3. Agent-assisted evidence review. ZeroDark’s custom harness coordinated AI agent swarms using locally run open-source and fine-tuned models in NVIDIA OpenShell sandboxes. The agents organized approved observations, identified duplicate evidence, and helped map findings to the selected profiles. Runtime guardrails limited access to approved material. The lead tester retained control of scope and validation.
  4. Human validation and reporting. The tester confirmed the VPN authentication behavior, portal rendering, staging exposure and TLS observations. The credential owner confirmed the exposed key’s status in the Claude Console. The tester recorded evidence and assigned severity based on exposure, prerequisites and business impact.
  5. Retest. The tester reviewed the original evidence against the client’s remediation changes and recorded a separate disposition for each finding.

Threat actor perspectives

These profiles were chosen to assess different business risks. The scenarios describe potential impact for the fictional client.

Opportunistic scanner
Selected as a baseline for indiscriminate internet exposure. The staging directory and legacy TLS support were visible without a client account (ZD-01, ZD-02).
Organized cybercrime · Storm-0539 (Atlas Lion)
Microsoft has documented phishing and gift card fraud by Storm-0539. We selected this profile because the fictional client issues gift cards and relies on remote access for staff. Credentials compromised through phishing could be enough for VPN entry because the gateway lacks a second factor (ZD-03). The support portal XSS could make phishing more credible by abusing trust in the company’s domain (ZD-04). The resulting risk is conditional on credential compromise and the privileges of the affected account. Microsoft threat intelligence.
State-sponsored actor · Volt Typhoon
Selected as a comparison for the importance of protecting public-facing edge devices. U.S. agencies describe Volt Typhoon targeting critical infrastructure. Here, the relevant question was whether the VPN authentication and transport controls would hold up against a targeted adversary (ZD-03, ZD-02). The fictional client’s selection of this profile is a testing assumption. Joint agency advisory AA24-038A.

Strategic recommendations

  1. Secure remote access first. Require phishing-resistant MFA across VPN sign-in paths. Assign ownership for exceptions and review whether remote roles can reach more of the network than they need.
  2. Treat leaked keys as incidents. Track revocation, replacement, usage review and billing review. Set workspace spend controls and alert on unexpected increases in model use.
  3. Prevent repeat exposure. Standardize output encoding and current TLS configurations, keep staging private, and make secure defaults part of release acceptance. Track all open findings by owner and retest date.

Issue trend

Fictional monthly history. Bars show findings still open at month-end or, for November, after the November 11 retest. Verified closures can include issues carried from the previous month.

  1. September 20261 of 53 new · 2 closed · 1 open

    Public debug page and expired test certificate closed. Verbose error responses carried forward.

  2. October 20260 of 52 new · 3 closed · 0 open

    Verbose errors closed, along with a newly exposed admin page and permissive cross-origin policy.

  3. November 20262 of 55 new · 3 closed · 2 open

    VPN MFA, key disablement and private staging verified. Reflected XSS and legacy TLS remain open.

Issue summary

Issue details

ZD-03 · Internet-facing VPN uses single-factor authentication

High

Evidence E-03. A designated test account established a VPN session with its password alone. Compromised staff credentials could grant the network access allowed by that account.

Recommendation. Require phishing-resistant MFA for every remote user, review exceptions, and validate network access restrictions independently.

Reference. CISA phishing-resistant MFA guidance

ZD-05 · Anthropic API key in public GitHub history

High

Evidence E-05. A historical commit contained a redacted Anthropic Claude API key. The client’s Claude Console administrator confirmed it was active in the workspace used for model inference. A third party could make model requests billed to the client and consume workspace capacity.

Recommendation. Disable the key, replace it through a secrets manager, review usage and costs for the exposure window, and enable repository secret scanning and push protection.

Reference. Claude API authentication and key management

ZD-04 · Reflected cross-site scripting on the support portal

Medium

Evidence E-04. The portal rendered input from a GET query parameter without appropriate output encoding. A harmless browser check confirmed script execution on the public support origin. This could undermine trust in pages served under the company’s domain.

Recommendation. Apply context-appropriate output encoding and add a regression test.

Reference. OWASP cross-site scripting prevention

ZD-01 · Public directory listing

Medium

Evidence E-01. The public staging site returned a directory index listing internal release filenames.

Recommendation. Restrict staging access and disable directory indexing.

Reference. MITRE CWE-548

ZD-02 · Legacy TLS accepted

Low

Evidence E-02. The VPN endpoint still negotiated TLS 1.0.

Recommendation. Disable TLS 1.0 and 1.1, and verify that supported clients can connect using TLS 1.2 or later.

Reference. RFC 8996

Retest update

On November 11, the tester repeated the checks for the five original findings and reviewed evidence that the exposed API key was disabled with the credential owner.

ZD-03 · Single-factor VPN Fix verified
Password-only sign-in was denied for the designated test account. The VPN session required phishing-resistant MFA.
ZD-05 · Exposed API key Fix verified
The client showed the exposed key as disabled in the Claude Console and confirmed replacement. Usage and billing review remained with the client’s incident owner.
ZD-04 · Reflected XSS Still open
The tester observed script execution in the affected support portal response during the retest. Apply output encoding and repeat the check.
ZD-01 · Public directory listing Fix verified
The original staging path returned HTTP 403 from the public internet. No directory index was visible on the tested host.
ZD-02 · Legacy TLS Still open
The VPN still negotiated TLS 1.0. Disable the legacy protocol and repeat the transport check.

Appendix

Severity and disposition

High indicates significant potential impact to access, data or spend. Medium indicates material exposure with additional prerequisites or limited impact. Low indicates a hardening gap with narrower direct impact. Fixed means the specific remediation was verified during the retest. Open means the original condition remained observable.

Evidence and limitations

Evidence in this public sample is summarized and fictional. The IP addresses use documentation ranges. The observations reflect the assessment windows and the approved inventory. Threat intelligence informs scenario selection. Findings in this report do not establish activity by a named threat actor or confirm a compromise.

Hypothetical scenario

This sample illustrates how ZeroDark presents a penetration test and retest. The engagement, systems, evidence, and findings are hypothetical and describe no actual assessment of Cluckin’ Bell or the domains shown.

Request an external pentest