Sicurezza AI · Attacco basato su AI

Cisco research demonstrates VLMs can be manipulated by imperceptible pixel perturbations embedding instructions in images (e.g., banners/previews), creating a covert channel that humans and simple filters miss.

Cisco research demonstrates VLMs can be manipulated by imperceptible pixel perturbations embedding instructions in images (e.g., banners/previews), creating a covert channel that humans and simple filters miss.

1 fonte · 7 mag

Cronologia

• 7 mag Segnalato da SecurityWeek

Fonti

• 7 mag SecurityWeek

  Attackers Could Exploit AI Vision Models Using Imperceptible Image Changes

  Cisco’s AI security researchers have analyzed ways to target vision-language models (VLMs) using pixel-level perturbation.

  originale

Part of the PlainSec briefing for 2026-05-07

Every edition of this story: Cisco research demonstrates VLMs can be manipulated by imperceptible pixel perturbations embedding instructions in images (e.g., banners/previews), creating a covert channel that humans and simple filters miss.