Executive brief
The AI Tensor Engine for ROCm (AITER), a library used for accelerating AI workloads on AMD hardware, contains a critical security flaw in its communication component. An unauthenticated attacker can send a specially crafted message to the system that allows them to take full control of the AI inference workers. This could lead to the theft of sensitive AI models, data manipulation, or a complete shutdown of the AI processing infrastructure.
Technical details
A remote code execution vulnerability exists in AITER versions up to and including 0.1.14 within the MessageQueue.recv() function in shm_broadcast.py. The vulnerability is caused by the use of the insecure 'pickle' library to deserialize data received from a ZeroMQ (ZMQ) SUB socket without any authentication, HMAC signing, or format validation. An attacker who can reach the writer XPUB endpoint on the cluster network, or who can supply a forged Handle with a malicious remote_subscribe_addr, can deliver a crafted pickle payload. This payload is executed with the privileges of the inference worker process across all remote reader workers. While the attack requires network access to the cluster's internal communication ports, it requires no prior authentication.
Affected products
- AMD AI Tensor Engine for ROCm (AITER) through 0.1.14
Timeline
- 2026-05-13: other: Vulnerability identified as a known issue in project newsletter draft
- 2026-06-01: disclosed: Vulnerability disclosed by VulnCheck
- 2026-06-01: advisory: CVE-2026-49121 published