CVE-2025-68131 describes an information disclosure vulnerability in the cbor2 library, specifically when a CBORDecoder instance is reused across multiple decode operations. Values marked with the shareable tag (28) persist in memory, allowing attacker-controlled messages to read sensitive data from previously decoded messages if the decoder is reused across trust boundaries. This affects Red Hat AI Inference Server 3.2.2 (cuda-ubi9), Red Hat Enterprise Linux AI, and Red Hat OpenShift AI products processing untrusted CBOR data. The vulnerability is rated Moderate by Red Hat and has a CVSS base score of 5.3 (Red Hat) and 7.5 (NVD). The vulnerability is linked to CWE-212 (Improper Removal of Sensitive Information Before Storage or Transfer). No patch or official fix is currently available. Mitigation requires operational controls to avoid decoder reuse across trust boundaries or to segregate sensitive data processing from untrusted input.

This vulnerability allows an attacker to read sensitive data from memory that was decoded in previous operations if the CBORDecoder instance is reused across trust boundaries. This leads to information disclosure of potentially sensitive application data. The impact is limited to confidentiality; integrity and availability are not affected. The severity is rated moderate to high depending on the environment, with Red Hat assigning a Moderate rating and NVD a higher score due to confidentiality impact.

No official patch or fix is currently available for this vulnerability. Red Hat advises that applications using the cbor2 library should avoid reusing CBORDecoder instances when processing data from different trust levels. If reuse is unavoidable, ensure that sensitive data is not processed by a decoder instance that will later handle untrusted input. This operational mitigation prevents attacker-controlled messages from accessing previously decoded sensitive information.