A test of GPT-4V found that messages placed inside images could shift the system from its assigned task to another one.
The study examines “goal hijacking via visual prompt injection,” in which an attacker places instructions inside an image. When a large vision-language model reads the image, those instructions may redirect it from its original task to a different one chosen by the attacker.
In tests described by the researchers, GPT-4V showed a 15.8% attack success rate. The analysis also found that success depended on the system’s ability to recognize characters in images and follow instructions.
The researchers proposed a visual prompt-injection method called goal hijacking via visual prompt injection. It is designed to redirect a large vision-language model from its original task to an alternative task specified by an attacker. Their quantitative analysis found that GPT-4V was vulnerable, with an attack success rate of 15.8%. The researchers also reported that successful attacks relied on the model’s character-recognition and instruction-following abilities.
Many AI systems interpret both written prompts and visual inputs. The findings indicate that instructions embedded in images can create a security risk by changing what such systems do, including in applications where visual information is used to guide actions or decisions. The researchers say mitigation strategies will be needed for secure deployment in sensitive systems.
This is a quantitative analysis of a proposed attack method against GPT-4V. The abstract does not describe the number or type of test images, the full testing procedure, how the attack success rate was calculated, or how the results compare with other vision-language models. It therefore does not establish how often the vulnerability occurs across systems or real-world applications. The study identifies character recognition and instruction following as factors linked to attack success, but the abstract does not provide further detail about their individual effects.
New Generation Computing · 2026 · DOI: 10.1007/s00354-026-00334-8
Authors: Subaru Kimura, Ryota Tanaka, Shumpei Miyawaki, Jun Suzuki, Keisuke Sakaguchi
Institutions: Tohoku University, NTT (Japan)
Researchers tested a programmable hardware design that imitates selected quantum-computing ideas to run matrix calculations in parallel. In representative benchmarks, the design reached up to 6.3 times the speed and 4.1 times the energy efficiency of conventional matrix co-processors and embedded graphics processors.
Researchers evaluated three AI systems on hundreds of molecules bound to a SARS-CoV-2 protein and in three prospective drug screens. The results suggest that combining AI-based predictions with traditional docking may help prioritize candidates, although no single approach worked best in every test.
Researchers developed a hardware-assisted system intended to detect message injection, spoofing, flooding and impersonation attacks on automotive networks. In tests using simulated traffic and a prototype automotive processor system, it reported an average detection time of 14 microseconds and less than 2 percent CPU overhead.
