Researchers have demonstrated a novel side-channel attack technique that exploits vulnerabilities in graphics processing units to steal sensitive data from other users sharing the same cloud infrastructure. The attack, which the team has named "PixelSteal," requires no special privileges and can be executed by any cloud tenant with access to a GPU instance on the same physical hardware as the target.

The attack works by carefully monitoring GPU performance counters and memory access patterns, which can leak information about computations being performed by other processes sharing the GPU. In their demonstrations, the researchers were able to recover cryptographic keys, reconstruct images being processed by other cloud tenants, and infer details about machine learning models being trained on shared hardware.

"Cloud computing's multi-tenancy model creates inherent risks when resources are shared between mutually untrusting parties," said researcher Dr. Carlos Mendes. "This attack shows that GPUs have not received the same level of side-channel scrutiny as CPUs, and there are real vulnerabilities waiting to be exploited."

The findings have been shared with major cloud providers, GPU manufacturers, and relevant standards bodies. Several providers have begun implementing countermeasures, including more aggressive GPU memory scrubbing between tenant allocations and improved isolation mechanisms.

Until comprehensive fixes are available, organizations handling highly sensitive data are advised to use dedicated, non-shared GPU instances when running security-critical workloads in cloud environments.