Stealth benefit of image semantics when text alone suffices

Ascertain whether incorporating image semantics within the Cross-modal Adversarial Multimodal Obfuscation (CAMO) prompts provides additional stealth benefits in scenarios where textual cues alone are sufficient to execute the attack.

Background

CAMO can succeed in both text-only and image+text configurations. When the textual component already enables reconstruction, it is unclear whether adding semantically aligned images improves evasion of safety filters or reduces detectability.

Clarifying this would inform when to include or omit visual content in practice and help shape defenses that leverage cross-modal consistency checks.