In a recent report published by Trail of Bits, researchers find malicious hidden prompts in language learning models (LLMs) specifically when processing visual media. These prompts, easily embedded into the system by hackers, are exploited to downscale and manipulate image prompts through unauthorized actions. While on the surface everything appears to be normal by users, AI carries out hidden instructions at the back end that pose serious risks for data privacy and personal identity. The researchers urge the public to be cautious and take preemptive measures such as firewall security to ensure their online safety.