AI Security · AI-Powered Attack

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 source · May 7

Timeline

May 7 Reported by SecurityWeek

Sources

May 7 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.
original

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.

AI Security · AI-Powered AttackCisco 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 source · May 7Timeline

May 7 Reported by SecurityWeek

Sources

May 7 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.

Part of the PlainSec briefing for 2026-05-07Every 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.