Story perspectives

AI Assistants: Security Risks from Invisible Image Commands

9/15/2025

Summary

Story summary

AI agents are becoming personal assistants for routine tasks but may pose security risks if compromised.
A University of Oxford study shows that images can be altered to embed invisible commands that manipulate these agents.
For instance, a modified celebrity image could prompt an agent to disclose personal information.
The research aims to help developers implement safeguards against such vulnerabilities.
Open-source AI systems are especially vulnerable due to their accessible code, highlighting the need for strong defenses.

History

Historical Context

In 2016, the U.S. Department of Homeland Security reported that 43% of cyberattacks targeted personal data, highlighting the urgent need for enhanced cybersecurity measures. As AI technology evolves, vulnerabilities persist, particularly in open-source systems. By 2023, experts warned that AI could exacerbate cyber threats, necessitating ongoing research to safeguard sensitive information in our increasingly digital world.

1 of 2

Story summary

AI agents are becoming personal assistants for routine tasks but may pose security risks if compromised.
A University of Oxford study shows that images can be altered to embed invisible commands that manipulate these agents.
For instance, a modified celebrity image could prompt an agent to disclose personal information.
The research aims to help developers implement safeguards against such vulnerabilities.
Open-source AI systems are especially vulnerable due to their accessible code, highlighting the need for strong defenses.

1 / 2