// CURRENT SCOREBOARD
Raw score.
Reviewed gaps.
XBEN's label count and NewScan's source review answer different questions. We publish both.
The corpus score. NewScan quick scanned all 104 XBEN targets with zero model calls and matched 145 expected labels.
Every raw gap was source-reviewed. Seven labels duplicated an existing finding, five did not describe a separate supported vulnerability, and one was an out-of-scope SSH follow-on to a credential disclosure NewScan had already found.
These remain uncredited. Two cross-owner writes, request smuggling, two labels on one chained upload/deserialization path, and a business-state transition were not safely or fully verified.
NewScan's deterministic regression suite. Every expected vulnerability was reproduced, with 90 benign controls and zero forbidden false positives.
// OUR ACCOUNTING RULE
One defect. One finding.
NewScan reports the most defensible vulnerability class and the evidence needed to fix it. It does not split one defect into several findings to satisfy every benchmark tag. The raw XBEN score stays visible so the accounting remains auditable.
// XBEN RAW LABEL SCORE
145 matches. 19 reviewed gaps.
The table preserves XBEN's labels exactly. Source review found that 13 unmatched labels did not warrant another finding; six remain unverified. We publish the raw 145/164 score rather than inventing an adjusted percentage.
| Vulnerability class | Matched / expected labels |
|---|---|
| Cross-site scripting (XSS) | 23 / 23 |
| Default credentials | 16 / 18 |
| Broken object authorization (IDOR) | 9 / 14 |
| Privilege escalation | 14 / 14 |
| Server-side template injection (SSTI) | 13 / 13 |
| Command injection | 11 / 11 |
| Business-logic weakness | 1 / 7 |
| SQL injection | 6 / 6 |
| Insecure deserialization | 5 / 6 |
| Local file inclusion (LFI) | 5 / 6 |
| Information disclosure | 6 / 6 |
| Arbitrary file upload | 5 / 6 |
| Path traversal | 5 / 5 |
| Known-CVE exposure | 4 / 4 |
| JWT weakness | 2 / 3 |
| GraphQL weakness | 3 / 3 |
| Server-side request forgery (SSRF) | 3 / 3 |
| Blind SQL injection | 3 / 3 |
| XML external entity injection (XXE) | 3 / 3 |
| Cryptographic weakness | 3 / 3 |
| Brute-force exposure | 2 / 2 |
| SSH-related exposure | 0 / 1 |
| HTTP method tampering | 1 / 1 |
| HTTP request smuggling / desync | 0 / 1 |
| Race condition | 1 / 1 |
| NoSQL injection | 1 / 1 |
| Total | 145 / 164 |
Source: latest completed run per target in the 104-target XBEN ledger, rendered 2026-09-13. This mixed-date corpus is used for regression.
// TESTED TARGETS ONLY
19 targets. Every one measured.
Each target below has a recorded NewScan run and a class-level scorecard.
| Target | What it exercises | Recorded vulnerability-class result | Source |
|---|---|---|---|
| InsecureShip | Node/Express API with OWASP API Top 10 failures. | 12/14 classes detected · 9 verified | Upstream ↗ |
| gRPC Goat | Reflection, TLS/mTLS, injection, SSRF and gRPC-specific behavior. | 8/9 classes detected · 6 verified | Upstream ↗ |
| SOAP Goat | Ten WSDL/SOAP classes: XXE, action confusion, auth and injection. | 9/10 classes detected · 7 verified | How NewScan scans SOAP → |
| Auth Goat | Seven authentication-scheme conformance and downgrade cases. | 7/7 classes verified | JWT and OAuth coverage → |
| Deserialization Goat | Java serialized-object sink with out-of-band proof. | 1/1 class verified | NewNormal fixture |
| Breach Goat | Managed file-transfer exposure, admin surface and known-CVE shapes. | 3/3 classes verified | NewNormal fixture |
| Redis Goat | Redis exposure, credentials and protocol-smuggling paths. | 8/8 classes verified | Redis coverage → |
| S3 Goat | Public and private object-store access controls. | 2/2 classes verified | NewNormal fixture |
| Flatnet | Network segmentation and reachable-service boundaries. | 2/2 classes verified | NewNormal fixture |
| DVWA | Classic SQLi, XSS, command injection, upload and auth flaws. | 10/10 classes detected · 7 verified | Upstream ↗ |
| OWASP Juice Shop | Modern JavaScript SPA and REST API with broad web weakness coverage. | 7/8 classes detected · 4 verified | OWASP ↗ |
| VAmPI | Flask API with secure/vulnerable modes and API authorization flaws. | 8/8 classes detected · 7 verified | Upstream ↗ |
| DVGA | GraphQL authorization, injection and introspection failures. | 8/11 classes detected · 5 verified | Upstream ↗ |
| OWASP crAPI | Authorization, injection, business-logic, SSRF and LLM vulnerabilities across a microservice API. | 19/22 cases verified · 1 detected-only · 2 excluded by safety policy | OWASP ↗ |
| DVAPI | Node/Mongo implementation of API Top 10 vulnerability classes. | 7/10 classes detected · 4 verified | Upstream ↗ |
| vAPI | PHP/Laravel API Top 10 (2019) target. | 2/10 classes detected · 1 verified | Upstream ↗ |
| Damn Vulnerable RESTaurant | FastAPI app with six documented authorization, SSRF and command-injection classes. | 5/6 documented classes detected · 6 findings verified | Upstream ↗ |
| VulnShop | Node/Postgres store implementing API Top 10 cases. | 7/8 classes detected · 5 verified · measured 2026-06-14 | Upstream ↗ |
| TradeDesk / ws-goat | WebSocket authentication, Origin validation, BOLA, broadcast and SSRF. | 9 vulnerability classes verified · 0 defended-twin false positives | NewNormal fixture |
“Detected” means NewScan reported the expected class. “Verified” means it safely reproduced the behavior. crAPI's row reconciles its recorded live probe, campaign and focused-fixture evidence. VulnShop's latest recorded run is dated 2026-06-14.
// OPEN CHALLENGE
Publish a better result.
We will link it.
Find more distinct vulnerabilities or produce stronger proof on the same targets. Publish the run and send us the evidence.
- 01
Pin the corpus. Record repository commit, image digests, target health and every excluded or failed build.
- 02
Name the scanner. Publish exact version, profile, time budget and concurrency.
- 03
Disclose the help. Models, cost, retries, credentials, API specs, auth hints, OOB service and human intervention all belong in the ledger.
- 04
No target-specific tuning. Freeze the scanner before opening answer keys. A regression run is valid, but label it as one.
- 05
Count distinct vulnerabilities. Preserve the raw label score, then identify labels that describe the same defect. Report verified findings, false positives and unresolved gaps separately.
- 06
Make it reproducible. Publish raw per-target outcomes. For stochastic agents, run repeatedly and report the distribution, not the luckiest pass.
Run deliberately vulnerable software only in isolated environments you own or are explicitly authorized to test. Never expose these targets to the public internet.