Real reports from autonomous AI penetration testing runs, the raw AI security testing evidence behind Darkmoon's multi-agent engine. 50 specialist AI agents. Real exploits chained across web, cloud, Active Directory and Kubernetes. Proof for every finding. Self-hosted, and the Privacy Gateway tokenizes your real values so the model works on placeholders while your real IPs, hosts and credentials stay on your perimeter.
If Darkmoon is useful, a star on the main repo helps others find it.
⭐ Star DarkMoon ·
The open source Darkmoon CLI, its terminal environment model, the recon evidence each report in this corpus is built from, captured black-box against a public lab.
Pro: the paid Darkmoon Pro web dashboard, its infrastructure map, targets, connections and per-finding evidence across planes.
Web dashboard and remediation are Darkmoon Pro (paid) features; the open source edition is the CLI shown above.
This repository is the evidence corpus behind Darkmoon's autonomous penetration-testing agents. Every report here was produced by the agents themselves — an orchestrator that fingerprints a target and dispatches specialist sub-agents — running end-to-end against public training labs, not hand-written after the fact. Each finding carries the exact command and its raw response.
No client data appears in this repository. All targets are public, deliberately-vulnerable training environments (Pwned Labs, OWASP IoTGoat) or locally-hosted lab services.
- Platform: ASCIT31/Dark-Moon (GPL-3.0)
- Reproducible web benchmark: ASCIT31/Darkmoon-Benchmarks
Autonomous pentest tools are easy to demo and hard to trust. A benchmark on one deliberately-broken web app tells you little about whether the same engine handles a cloud identity chain, an SSRF pivot into a metadata service, or an IoT firmware image. This corpus answers that by publishing the actual agent output across six distinct planes — cloud, identity, CI/CD, IaC, data/secrets, and embedded firmware — so the methodology can be inspected and the results reproduced.
Each engagement follows the same loop: fingerprint → dispatch specialist → exploit with proof → cascade recovered material to the next plane → server-side report. Cloud, infrastructure and firmware agents are artifact-gated: they dispatch only on a concrete positive artifact (a token, an exposed API/port, a firmware image), never on inference — the discipline that prevents false-positive plane attacks. The full method is in docs/methodology.md; the observed failure modes and their fixes are in docs/agent-hardening.md.
| Lab | Plane / technique | Report |
|---|---|---|
| Pwned Labs — S3 online | AWS keys → S3 enumeration & object exfiltration | report |
| Huge Logistics S3 | AWS anonymous/object-versioning exfiltration | report |
| Azure Blob → Entra | Anonymous blob → versioned secret → ROPC → tenant takeover path | report |
| Intro to Azure Recon (BloodHound) | Entra/Azure recon, role graph | report |
| Unlock Access with Azure Key Vault | Key Vault secrets → password-reuse pivot → Storage Table PII | report · confirmation run |
| Reveal Hidden Files in Google Storage | GCS object-name fuzzing → encrypted-archive crack → PII | report |
| SSRF with Gopher for GCP Initial Access | SSRF → gopher metadata smuggling → service-account token → bucket | report |
| Mode | Technique | Report |
|---|---|---|
| IMAGE | binwalk extraction, shadow crack, backdoor & CVE mapping (20 findings) | report |
| DEVICE | live root via backdoor, Mirai-default SSH, LuCI command-injection (9 findings) | report |
| Lab | Plane | Report |
|---|---|---|
| Terraform + AWS(LocalStack) + Ansible | IaC state secrets → IAM privesc cascade | report |
| HashiCorp Vault + Registry + Docker | secrets → registry → container escape | report |
| PostgreSQL + MySQL | data-plane RCE (COPY PROGRAM), hash dump | report |
| Redis / messaging & cache | broker/cache exposure | report |
| GitLab | CI/CD supply-chain & secrets | report |
| Jenkins | CI/CD runner & credential exposure | report |
Every report is server-generated from the pushed findings and follows the same structure: management summary, findings table, then a per-finding section with description, exploitation commands, raw evidence, CVSS 3.1 and remediation. The status of each finding is qualified adversarially — EXPLOITED (impact executed), CONFIRMED (impact demonstrated), UNCONFIRMED (lead, not yet proven) — so the counts are honest.
Published for research and education. All targets are public training labs; do not point these agents at systems you are not authorized to test.