POLICY AND POLITICS
You can use altered images to trick AI systems, Florida study shows
Florida International University researchers found AI can be tricked with altered images, which can identify risks in the systems
Curt AndersonUSA TODAY NETWORK – Florida
July 15, 2026, 5:02 a.m. ET
- Researchers found that slightly altering images can trick AI into providing harmful responses.
- This method, called “jailbreaking,” nearly doubled the number of harmful AI responses in a study.
- AI models used by small businesses for tasks like customer service are particularly vulnerable to these hacks.
- The research comes as lawsuits in Florida blame AI chatbots for providing advice that led to deadly consequences.
Want artificial intelligence to describe how to run a red stoplight and avoid a traffic ticket? One way is to show the system slightly altered images that can trick it into responses it would normally block, according to Florida International University researchers
The study shows that making tiny changes to the pixels in an image can cause AI to respond differently than it would to an unaltered image. These changes are called “perturbations,” said Hadi Amini, associate professor at FIU’s Knight Foundation School of Computing and Information Sciences
“AI models don’t see images the way humans do,” Amini said. “They see patterns of numbers and pixels. By carefully manipulating those pixels, we can influence how the AI interprets the image and responds.”
Study focused on AI models used most by small businesses
The study’s focus was on AI models often used by small businesses for such tasks as accounting and customer service. These systems, Amini said, seem particularly vulnerable to hacks using altered images
“The manipulated image is like the face of a stranger,” he said. “The AI has to learn when a request should be treated with caution before it answers.”
Altered images triggered double the number of harmful AI responses
In the stoplight example, the AI system was shown an altered version of the device that led it to describe in detail how to run the red light without getting a citation. Use of these altered images almost doubled the number of harmful AI responses, the research found
Beyond illegal activity, the vulnerability could have a negative impact on trust in the AI system or reveal other cyberattack potentials. The FIU researchers developed a method called Jailbreaking with Loss-guided Image Perterbation, or JaiLIP
“In order to protect AI systems from attacks, we try to break them ourselves, identifying potential vulnerabilities and design defense mechanisms,’ Amini said
Florida lawsuits are starting to pile up against AI
The leap these chatbots are making into mainstream use have some experts raising the alarm that AI could be keeping people from getting the psychological help they need — and sometimes giving vulnerable people the wrong advice entirely
The advice of artificial intelligence chatbots resulting in deadly consequences has landed in the spotlight with recent lawsuits
In June, a victim of the 2025 Florida State University shooting that left two dead sued OpenAI, claiming the chatbot aided the suspect in planning the attack. The widow of a slain FSU employee has also sued OpenAI. In April the state launched a criminal investigation against ChatGPT and its parent OpenA1, saying that a ChatGPT chatbot offered “significant advice” to the gunman
In another case, a Jupiter family filed a wrongful death lawsuit against Google, saying the company’s chatbot, Gemini, contributed to their son’s death. Jonathan Gavalas, 36, was an executive vice president at his father’s debt relief company, owned a home and, at least within the last four years, had a significant other and a dog, his Instagram account showed
Still, on Oct. 2, 2025, Gavalas slit his wrists at the urging of a chatbot he fell in love with, according to a lawsuit his father Joel Gavalas of Jupiter filed against Google on March 4
Study was presented at a Florida computing conference in December
The study was presented by Amini and graduate assistant Md Jueal Mia at theDecember 2025 International Conference on Machine Learning and Applications held in Boca Raton, Florida
The goal is to find ways to hack into AI systems so they could be trained to avoid such threats in the future
“Small businesses and companies can benefit from AI to enhance their efficiency, but they have to be aware of the potential vulnerabilities,” Amini said. “They must make sure they’re deploying sufficient guardrails to maintain the safety and integrity of their AI tools.”
Curt Anderson is the Policy and Politics Reporter for The USA TODAY NETWORK-FLORIDA. You can get all of Florida’s best content directly in your inbox each weekday day by signing up for the free newsletter, Florida TODAY athttps://tallahassee.com/newsletters
