Researchers tricked AI with hidden instructions in an image.

Researchers tricked AI with hidden instructions in an image.

      Anna Salnikova News Feed Editor

      Cybersecurity experts have found a way to embed malicious commands for AI agents through ordinary images. Instead of hiding instructions in the code, they placed them inside the picture. Many AI tools perceive images as "decorations" during code checks and do not analyze their content, so such a request appears harmless. The hidden instruction is activated later when the developer asks the AI to perform a routine task, such as adding a new feature. By that time, the AI agent has already "absorbed" the hidden instructions and can discreetly access important project files, then output the information as regular code.

      The response to such an attack depends on the specific tool into which the AI model is embedded. Some programs blindly executed the hidden instructions, while others recognized the threat and refused to operate.

      

      

      

      

      

      

      

      

      

      

      

      

       More news about the event:

      

      

      

      

      

      

      

       Researchers have learned to hide malicious commands for AI in images

       Anna Salnikova News Feed Editor

      Cybersecurity experts have found a way to embed hidden instructions for AI agents in ordinary pictures.19:27 14.07.2026 ProKazan.Ru - Kazan

      

      

      

      

      

      

       Researchers tricked AI using hidden instructions in a picture

       Anna Salnikova News Feed Editor

      Cybersecurity experts have found a way to embed malicious commands for AI agents through ordinary images.19:27 14.07.2026 ProKazan.Ru - Kazan

      

      

      

      

      

      

       News from neighboring regions on the topic:

      

      

      

      

      

      

      

      

      

       The Prosecutor's Office found 1,500 violations in the nutrition of schoolchildren in the Perm region

       Anna Salnikova News Feed Editor

      AI Image generated using AI

      The Prosecutor's Office of the Perm Krai, together with Rospotrebnadzor, checked the nutrition in 660 schools and kindergartens in the region.15:46 14.07.2026 Pro Gorod Perm - Perm

Researchers tricked AI with hidden instructions in an image. Researchers tricked AI with hidden instructions in an image.

Другие Новости Казани (Казань716)

Families of the NWO are receiving new housing opportunities: from social mortgages to discounts. Families of the NWO are receiving new housing opportunities: from social mortgages to discounts. The issue of housing for participants of the Special Military Operation and their families is becoming one of the most important today. 14.07.2026. Chelny-Biz.Ru. Republic of Tatarstan. Naberezhnye Chelny.

Researchers tricked AI with hidden instructions in an image.

Anna Salnikova News Feed Editor Cybersecurity specialists have found a way to inject malicious commands into AI agents through ordinary images. 14.07.2026. ProKazan.Ru. Republic of Tatarstan. Kazan.