Analiza pull requests de GitHub detectando vectores de supply chain, lifecycle hooks, ejecución remota y secretos expuestos. Los escaneos se ejecutan bajo demanda o vía integración con GitHub; Sentinel no reescribe los resultados de GitHub ni declara limpio lo que no ha inspeccionado.
These offline evaluations cover different tasks and cannot be combined into a single ranking. The NPM result has a small benign sample and high false-positive rate; the synthetic PR set is not evidence of generalization. No result here establishes general superiority over another security tool.
| Evaluation Metric | Sentinel Cloud | Sentinel CLI | Sentinel Oracle | Scope and limitation |
|---|---|---|---|---|
| OWASP Benchmark v1.2 — Java (2,740 cases) | TP 1,336 · FP 1,259 · FN 79 · TN 66 · Score 0 | Not evaluated | Not evaluated | Cloud recall 94.4%; FPR 95.0%. This corpus does not establish Java SAST capability. |
| NPM malicious annotated-file recall (602 samples) | 87.9% (529 / 602) | 83.1% (500 / 602) | 87.5% (527 / 602) | Cloud / CLI / Oracle respectively. Sentinel scanned annotated malicious files, not full packages; this comparison is optimistic and not apples-to-apples. |
| NPM benign sample false-positive rate (32 packages) | 87.5% (28 / 32) | 65.6% (21 / 32) | 81.3% (26 / 32) | Cloud / CLI / Oracle respectively; the curated sample is small and is not the paper’s full benign corpus. |
| Synthetic PR-level malicious recall (60 malicious / 40 benign) | Cloud 76.7% (46 / 60) | CLI 65.0% (39 / 60) | Oracle 15.0% BLOCK (9 / 60) | Benign PR false-positive rates: Cloud 20%, CLI 10%, Oracle 0%. Oracle BLOCK+REVIEW is 46.7%; these are distinct decision thresholds. |
| Determinism (100 synthetic PRs × 2 passes) | Cloud 100% | CLI 100% | Oracle 100% | Observed agreement on this corpus only; it is not evidence of detection quality or generalization. |
| LLM usage in the recorded offline runs | 0 calls | 0 calls | 0 calls | These offline runs do not measure every production Cloud operation. |
Raw counts are preserved in docs/benchmarks/results/owasp-2026-08-16T02-19-43-975Z.json. FPR is FP / (FP + TN) = 1,259 / 1,325 = 95.019%; an older summary said approximately 93%. The historical JSON is unchanged.
TP means a labeled vulnerable case was flagged; FP means a labeled benign case was flagged; FN means a labeled vulnerable case was missed; TN means a labeled benign case passed. Recall = TP / (TP + FN), FPR = FP / (FP + TN), precision = TP / (TP + FP). OWASP labels its Java cases; the NPM set uses paper annotations; the PR set uses synthetic labels. None of these results is a paired vulnerable-versus-patched CVE test.
References: OWASP Benchmark · NPM Malicious Package Study. Historical engine commits: Cloud worker `5b5e1dd`, CLI `bb95341`, Oracle `786b55f`; runs were offline/local, with Oracle tarball and content-intelligence integrations disabled.
The Cloud engine run against real, well-known open-source repositories. Raw signals are engine output; human-verified are the findings that survived manual triage and map to a shipped change. Every false positive is documented, not dropped.
| Repository | Scope | Raw | Verified | Outcome |
|---|---|---|---|---|
| axios/axios | workflows + lib/ source | 59 | 0 (documented FPs) | CLEAN |
| nodejs/undici | npm publish pipeline | 31 | 1 — provenance gap | PR #5880 |
| pnpm/pnpm + pnpr registry | CI + 11 Rust crates | 204 | 0 (documented FPs) | CLEAN |
| vitejs/vite | .github workflows | 44 | 0 (documented FPs) | CLEAN |
| fastify/fastify | package-manager CI | 22 | 1 — curl|bash without pin | PR #7050 |
| drizzle-team/drizzle-orm | release workflows | 68 | 0 (placeholder creds) | CLEAN |
| modelcontextprotocol/typescript-sdk | 8 workflows | 29 | 29 — mutable tags | PR #2603 |
| punkpeye/fastmcp | CI/CD workflows | 7 | 7 — hardening gaps | PR #303 |
Raw = engine findings pre-triage · verified = survive human verification, map to a change · PRs are supply-chain hardening framing
Sentinel Cloud analyzes pull-request changes and presents its own findings with repository and scan context. The output is evidence for a reviewer and may vary when the scanned commit, enabled checks, or integrations change.
GitHub check runs and Code Scanning alerts are shown in a separate evidence view. They are not folded into Sentinel findings or used to rewrite the Sentinel scan result.
So we do not replace your AI. We feed it. Cloud returns deliverables as structured JSON — verdicts, kill chains, evidence — designed to be consumed by both humans and AI tools. Your agent stops re-reading raw diffs and starts answering from pre-filtered evidence.
These products are not assigned comparative scores here. The descriptions below summarize the integration boundaries documented in their current contracts.
Capturas operativas reales del motor Sentinel Cloud analizando repositorios de alta exposición, trazando grafos de kill chain y generando recibos criptográficos deterministas.
Monitor centralizado con el estado de sincronización de repositorios, métricas de escaneo y protección activa en Sentinel Cloud.
Clasificación precisa de hallazgos por severidad (CRITICAL, HIGH, MEDIUM, LOW) con veredicto final determinista.
Tarjeta forense detallando la carga útil identificada, fragmento de código, categoría de amenaza y mitigación sugerida.
Correlación multietapa que mapea cómo una instrucción secundaria se conecta a un vector de exfiltración o ejecución remota.
Prueba de auditoría inmutable firmada digitalmente para garantizar verificabilidad sin drift en el pipeline CI/CD.
Evaluación cuantitativa frente a marcos estándar de la industria (OWASP, NPM Malicious Datasets), mostrando recall y precisión.
Six layers, one pipeline. Each layer feeds the next; kill chains are only built when signals across layers correlate. The worker path is what production runs on every PR.
Every number on this page comes from a benchmark you can run yourself against the real production scanner — not from a vendor-commissioned report. If a claim breaks, the fix lands in the open test corpus, not in a footnote.
The two retention numbers on this site are different by design: 365 days is the Cloud metadata retention (ENTERPRISE); 14 days is the internal offline calibration buffer (hash-only, never customer data). Neither ever stores your source files.