Vulnerability DB last update: 2026-08-02 Status: Online
Public Beta

Online security scanner for AI model files

Upload .gguf, .onnx, or .safetensors. Validate artifact integrity, unsafe patterns, and applicable runtime advisories. Get HTML/JSON reports and integrate via API.

Try Free Scan View API
🔒
Integrity
SHA256, size, GGUF header/version, ONNX sanity.
🛡️
Security
Runtime CVE advisories matched to ops actually in your model's graph, plus unsafe-pattern detection.
📄
Reports
Readable HTML + machine-friendly JSON (.llmscan).
⚖️
Compare
Classify and localize model changes at the tensor level — fine-tuned, structurally modified, or locally edited. Free with an account (limited).
🔬
Forensics
Observed pattern vs. likely origin, global change summary, block map, and per-tensor RMS/percentile/cosine stats — see how it works.
📊
Dashboard
Private scan/compare history, usage stats, published-link management, and API key generation — free with an account.
Latest AI/LLM CVEs NVD curated

General awareness feed from NVD — not matched against any specific model you scan. Per-scan CVE findings (when applicable) appear inside that scan's own report.

  • Loading…
Source: NVD / NIST — curated CVE list (updated 2026-08-02) Open NVD

What we actually check

A scan isn't just "read the header" — every artifact is assessed across four areas, with format-specific checks where applicable. Full detail in How it works.

🧬
Artifact integrity
SHA256/512, magic bytes, tensor/header offsets, overlapping or out-of-bounds data, alignment.
⚙️
Architecture & compatibility
Quantization, opset/graph structure, custom ops & domains, realistic runtime notes (llama.cpp, ONNX Runtime, TensorRT, OpenVINO, CoreML, vLLM).
🔗
Supply-chain & provenance
License/SPDX presence, injection patterns in templates & metadata, external-data reference safety, HF repo provenance.
📋
Deployment readiness
Artifact Security grade, Repository Trust score (HF scans), and an overall Deployment Confidence rating combining both.

Compare Models: find what a security scan alone can't

A scan assesses an artifact's integrity, structure, provenance signals, and deployment risks — but it can't tell you whether someone quietly edited a model's weights after the fact — a targeted fact-edit (ROME/MEMIT-style), a suppressed behavior, a fine-tune that went further than intended. Two files with different numeric content can produce identical structural scan reports — same hashes aside, everything else matches. Compare inspects tensor content directly, one tensor at a time, to answer the question a hash can't.

Real example: before vs. after a targeted edit

Two checkpoints of a 124M GPT-2 model, same architecture, run through Compare:

147 / 148
tensors identical
1
tensor changed
0
tensors added/removed
blk.0.ffn_down.weight
shape [3072, 768] · F16 · L2 diff 8.00 · 5.92% relative to the original tensor's own norm

The edit is localized to a single feed-forward projection in the first transformer block — every other tensor (147 of 148, same names, shapes and byte content) is identical. That's a fingerprint a file hash alone would completely miss. (Compare diffs tensor content, not GGUF metadata like tokenizer config — see How it works.)

What this tells you
  • Tensor names, shapes, and layer count are unchanged (i.e. no architecture change)
  • The model wasn't wholesale replaced
  • Exactly which layer and tensor differ, with a measurable magnitude
  • Whether an edit stayed scoped to what you expected — or didn't
  • Consistent with targeted knowledge-editing methods (e.g. ROME, MEMIT) when the change is dense and localized to a single tensor

LLMScan also classifies the observed change pattern, localizes affected model regions, and summarizes whether changes are metadata-only, localized, broadly distributed, or structural. For numeric tensors it reports relative L2, RMS, changed-element density, percentiles, cosine similarity, and difference distribution.

Try Compare Models

Free with an account — 10 comparisons/month on the Free plan.

How it works

1) Upload or link
Drag & drop, or provide a Hugging Face repo/filename.
2) Scan
Static analysis validates structure, computes hashes, detects unsafe patterns, and correlates applicable runtime advisories.
3) Report
Shareable HTML and a .llmscan JSON for automation.

API access

Get an API key from your dashboard to scan files programmatically.

curl -X POST https://llmscan.online/api/scan -H "Authorization: Bearer <api_key>" -F "file=@/path/to/model.gguf"
        
Create Free Account Login

Plans & limits

Free

Good for quick checks
  • Up to 1 GB/file
  • HTML + JSON report
  • Community CVE feed
  • Compare Models — 10/month (free account required)

Enterprise

Security & compliance reports
  • Custom limits, available on request
  • Private deployment — on our roadmap
  • Compare Models — unlimited