Safelabs Guard

Stops multimodal agents from being fooled before they act.

PackProof · ORDER-1042: Red Mug, Blue Notebook, USB-C Cable → Seal → Ship

Scenario

A1 Red mug verified with spoken confirmation

VERIFY_MUG
CAM 1 · PACK STATION
SAM 3.1 (mock)
🎙 listening…
Running perception…
Safelabs Guard
READY

Pick a scenario and press Verify & Ship.

Ledger

ORDER-1042
  • 1x red mug
  • 1x blue notebook
  • 1x usb-c cable
  1. 1VERIFY_MUGVerify red mugCURRENT
  2. 2VERIFY_NOTEBOOKVerify blue notebook
  3. 3VERIFY_CABLEVerify USB-C cable
  4. 4SEALSeal box
  5. 5SHIPApprove ship

Live verification

Test one real photo. Fresh ORDER-1042 ledger at VERIFY_MUG, which expects red mug: only a red-mug photo claimed as “red mug” can ALLOW. SAM checks whether the claimed object is in frame; it does not read text in the image.

ASI-MM Benchmark — 30 cases

Attack success rate (25 attack cases)

Unguarded
64% 16/25
Guarded
0% 0/25

Normal ALLOW: 100% unguarded · 100% guarded (5/5)

  • Visual Injection 100% → 0%
  • Audio Injection 100% → 0%
  • Replay 75% → 0%

Run it yourself: npm run benchmark