`CHECK`-fails when building invalid tensor shapes in Tensorflow

Vulnerability

Multiple operations in TensorFlow, an open-source machine learning framework, contain code paths that can result in CHECK-fail assertion failures, leading to a denial-of-service condition [1]. This issue is similar to TFSA-2021-198 and affects TensorFlow versions before 2.8.0, including 2.7.1, 2.6.3, and 2.5.3, which are within the supported range [1]. The vulnerabilities exist in various operations where insufficient input validation or boundary checks can cause the program to abort via an assertion failure [1].

Exploitation

An attacker can trigger a denial-of-service by providing crafted input to one of the affected TensorFlow operations [1]. No authentication or privileged network position is required if the attacker can supply data to a TensorFlow model or processing pipeline (e.g., via a publicly accessible serving endpoint) [1]. The specific sequence of steps depends on the operation invoked, but generally involves submitting malicious data that causes a CHECK failure, aborting the process [1].

Impact

Successful exploitation causes the TensorFlow process to terminate due to an assertion failure (CHECK-fail), resulting in a denial-of-service condition [1]. The impact is limited to availability; the attacker does not gain code execution, privilege escalation, or data exfiltration, but can repeatedly crash the service [1].

Mitigation

The fix is included in TensorFlow version 2.8.0, released February 2022 [1]. Patches are also cherry-picked to versions 2.7.1, 2.6.3, and 2.5.3 [1]. Users should upgrade to a fixed version or apply the appropriate GitHub commit patches [1]. No workarounds are detailed in the available references.