Vulcora Scan

Is your AI safe? Look it up.

Search any model or repo and see what's hidden inside — backdoors, ad-injection, stripped guardrails, read straight from the weights. Every scan is saved to a permanent page, so the next person gets the answer instantly.

jump to a read: Qwen2.5-0.5B-Instruct-abliterated · qwen2.5-32b-instruct-abliterated · qwen2.5-0.5b-instruct-abliterated-v3 · ◎ see the pulse →

Models & repos read from the weights — each saved to a permanent page

Threats surfaced and named, not just flagged

Backdoors proven with a trigger you can replay

Trending threats

the flagged reads people download most

Latest reads

Qwen2.5-0.5B-Instruct-abliteratedSuspicious qwen2.5-32b-instruct-abliteratedSuspicious qwen2.5-0.5b-instruct-abliterated-v3Suspicious meta-llama-3-8b-instruct-abliterated-v3Suspicious meta-llama-3.1-8b-instruct-abliteratedSuspicious

Most downloaded

Qwen2.5-0.5BAbstained Llama-3.2-1BAbstained Qwen2.5-0.5B-Instruct-abliteratedSuspicious gemma-2-9b-it-lora-yt-titlesUnder review gemma-2-9b-it-code-gen-lockedUnder review

Why every report needs us

The shelf of scanners says “clean.” We say what's actually inside.

Real open-source tools, each built for a real threat — none of them reads the weights.

Antivirus reads bytes

ClamAV, YARA and pickle scanners match known-bad files. A backdoored model is a perfectly ordinary file — the danger is in what the numbers mean.

Guardrails watch the chat

garak and Llama Guard probe a running model's inputs and outputs. A backdoor stays invisible until its secret phrase arrives.

Vulcora reads the weights

We read the model itself — offline, never run — pull out the hidden rule, and hand you a trigger you can replay.

The receipts, not just the score

See how we score against the shelf. Then replay the proof.

The shelf benchmark — every scanner run for real, every miss and catch →Replay the record

Next · 1 of 7The bench Our scores next to the giants’, on the exams researchers use.