American researchers discovered A vulnerability exists in GPUs from Apple, Intel, AMD, Qualcomm, Arm and NVIDIA that allows websites with malicious code to steal data from other websites – this could include logins, passwords and other sensitive information displayed.
The attack type based on the discovered vulnerability is called GPU.zip – it allows you to use GPU resources to steal data from websites whose pages are inserted into other malicious resources via iframe elements. Researchers have found that data compression performed by integrated and discrete graphics cards to improve performance can be a side channel for information theft on a pixel-by-pixel basis.
Modern GPUs automatically attempt to compress this visual data without involving the software that processes this data – this is done to save memory resources and improve performance. Notably, the attack works in Chrome or Edge browsers, but not in Firefox and Safari. The researchers took the Wikipedia website as a conditional victim, the code of which was embedded on the page of their own resource via an iframe. The attack worked on GPUs from Apple, Intel, AMD, Qualcomm, Arm and NVIDIA. On the AMD Ryzen 7 4800U, it took about 30 minutes to render the target pixels with 97% accuracy. With the Intel Core i7-8700 the time increased to 215 minutes.
All GPUs use their own data compression methods when outputting content – these are not documented, but scientists have been able to reverse engineer each. To do this, they used the SVG vector graphics format and captured the differences between black and white pixels in DRAM traffic. The work describes a method for data theft through integrated graphics, but a similar method is also implemented through discrete graphics.
To carry out an attack in a browser, three conditions must be met:
- Loading iframes with cookies is allowed;
- Allowed display of SVG filters in iframes;
- Rendering is done by the GPU.
In response to the query, a Google representative stated that a webmaster can protect his resources from attacks using the HTTP headers X-Frame-Options and Content Security Policy. In addition, the attack is technically quite difficult to implement and takes a long time, which reduces the level of threat to ordinary users. Intel and Qualcomm said the vulnerability associated with GPU.zip does not apply to graphics processors but to third-party software, so the companies will not take any action. Apple, NVIDIA, AMD and Arm did not comment.
While the GPU.zip attack doesn’t pose a significant threat to the average user, it could be the basis for other, more dangerous exploits, according to the researchers who discovered it. In addition, it should be noted once again that hardware optimization can create side channels of data theft that cannot be defused by software.