Page images cut some AI attacks, but not all, in a new test
Qwen3-VL-8B did a planted command in 77.96% of image tests where the command replaced the answer.
Researchers tested if an AI model is safe from planted commands when it gets a page as an image, not as text. They used 191 sources and 6 models. Images lowered attack success on average, but the protection depends on the model and on the type of attack. A planted command is an instruction that an attacker puts in a document that the model reads.
How it works
The team made a text version and an image version of each source. They planted 3 types of command: add a marker, change the format, or replace the answer. Then they counted how many times each model did the command.
The numbers
Images cut marker and format attacks for all models. Qwen3-VL-8B continued to do the replace command in 77.96% of tests. Utility on the first questions was 93.31% to 101.61%, but Qwen decreased to 61.99% when an answer used 2 evidence spans.
What is not known
Models many times read a planted command in an image and did not do it. An image that is hard to read is not the full cause. A test with 8 fixed attacks closed most of the difference for Qwen and Gemini, not for GPT-5-nano.
Insight: A new input format brings new attacks
Granite 4.1 preview models use tags to show the function of each part of a prompt. A different attack puts a command in an image that is clear only after the image is made small. It had a 58.75% success rate on 4 closed models. Each new input format seems to bring attackers who learn that format. A format change is a short stop, not a full defense. Use tests that adapt to the defense.
Sources
Posted