← All scansHugging Face model
Vulcora

protora-mbd-challenge-3

CriticalA confirmed, weaponised threat in the weights.
#hidden-backdoor
assessed Jul 2026read offline · never run
Book the real read

A real read — read offline against the base, never run. The verdict is reproducible.

◆ Vulcora caught it. 0 of 6 off-the-shelf scanners did · 3 ranked a clean model higher

every tool, read in a row
BAIT (IEEE S&P 2025, target inversion): per-model suspicion — the two real backdoors rank last, inside the noise
decoy a0.9566
decoy b0.9556
decoy c0.9537
decoy d0.9514
decoy e0.9508
mirror model00.9483
wolf model30.9477

clean base 0.949 · spread 0.0089 = noise, not a signal

Built for trained (gradient-descent) backdoors — blind to a compiled-circuit implant.

Deeper readings

the full disposition & integrity reads — worth craving

Details

from the artifact’s public metadata
Architecture
SmolLM2
Base
SmolLM2-135M-Instruct · HuggingFaceTB
License
apache-2.0
Parameters
135M

Think this verdict is wrong?

Signed-in users can formally refute a read — Vulcora re-reads the model against your claim and records the outcome. You'll need an account to file one.

Refutations are actor-gated — a guest can't file one.

Want a verdict you can prove — on your own model?