LLMScan — Frequently Asked Questions

What file formats do you support?

GGUF, ONNX and Safetensors, each with format-specific deep static analysis (structure/header validation, injection-pattern detection, and — where applicable — supply-chain and runtime-compatibility checks; depth varies by format, see How it works). Works both for direct file upload and for scanning a file straight from a Hugging Face repo URL.

What risks are detected?

We perform static checks (headers, sizes, hashes) and flag suspicious patterns (e.g., unsafe external data in ONNX) — we don't claim exploitability without further validation.

What does "CVE Advisory" in a report actually mean?

It's a runtime/parser advisory, not a claim that the model file itself is compromised. Concretely: for ONNX models, we check whether the operators actually present in your model's graph are among those affected by a known onnxruntime CVE, and tell you the minimum patched runtime version. It answers "is loading this graph with an old runtime risky", not "does this file contain malware" — those are different questions, and we only currently match one CVE category (runtime advisories for ONNX Runtime). The general "Latest AI/LLM CVEs" feed shown on the homepage is unrelated background awareness content (sourced from NVD) — it isn't matched against any file you scan.

Do you upload files to your servers?

Yes, for processing the scan — but only temporarily. See Security & Privacy for retention and isolation details.

How long do you retain files?

The uploaded model file itself is deleted automatically within 24 hours (or sooner) after scanning — we only keep the file long enough to run the scan. Operational logs exclude model contents.

Is my scan report public?

Scans run without an account are public demos: visible on the homepage's "Last free reports" list and reachable by anyone with the report link, retained for 30 days and then deleted automatically. If you need a private report, create a free account — reports from a logged-in account aren't listed publicly, and you can delete any of your own reports at any time from Scan History or Compare History (the entry stays in your history list — and counts toward your lifetime total — but the report content is removed, shown as "not available"). Once a deleted entry is more than 90 days old, the row itself is removed from your history list entirely; your lifetime scan/comparison counts on the Overview page are unaffected by this cleanup. Private reports you haven't deleted are kept indefinitely, the same across Free, Pro, and Enterprise — there's no plan-based expiry, only the 30-day window for anonymous demo scans and the 90-day purge for entries you've explicitly deleted.

Mobile models?

Supported if you can provide the raw model file (e.g., extracted from device or build artifacts).

How is this different from jailbreak/endpoint testing?

LLMScan analyzes the file (GGUF/ONNX), not the model’s conversational behaviour behind an API endpoint.

Can I run it locally?

Not on the hosted plans — but a Private Enterprise Deployment (a dedicated instance running inside your own network) is on our roadmap. Get in touch if that's what your team needs.

Limitations

Static scanning can’t guarantee absence of risk. Treat results as a security signal, not a formal audit.

False positives / contact

Report issues via contact@llmscan.online.