CVE-2025-65887 is a medium-severity vulnerability identified in the OneFlow machine learning framework, specifically version 0.9.0. The flaw exists in the flow.floor_divide() function, which performs floor division operations on input tensors. When this function receives a crafted input tensor containing a zero value as the divisor, it triggers a division-by-zero error, causing the application to crash and resulting in a Denial of Service (DoS). This vulnerability is classified under CWE-639 (Exposure to Incorrectly Controlled Resource Consumption), indicating improper handling of exceptional arithmetic conditions. The attack vector is network-based (AV:N), requiring no privileges (PR:N) but does require user interaction (UI:R) to supply the malicious input tensor. The scope is unchanged (S:U), and the impact is limited to availability (A:H), with no confidentiality or integrity impact. No patches or fixes are currently linked, and no known exploits have been observed in the wild. The vulnerability highlights the importance of input validation and error handling in numerical computations within AI frameworks. Given OneFlow's role in distributed machine learning workloads, exploitation could disrupt AI services or training pipelines.

The primary impact of CVE-2025-65887 is a Denial of Service condition affecting systems running OneFlow v0.9.0. This can interrupt machine learning model training or inference tasks, leading to downtime and potential loss of productivity. Organizations relying on OneFlow for critical AI workloads may experience service outages, delayed project timelines, and increased operational costs. Although the vulnerability does not expose sensitive data or allow code execution, the disruption of AI services can have cascading effects in sectors like healthcare, finance, autonomous systems, and research. The lack of required privileges means attackers can exploit this remotely if user interaction is possible, increasing the attack surface. The absence of known exploits reduces immediate risk but does not eliminate the potential for future attacks. Overall, the impact is moderate but significant for environments where availability of AI services is critical.

To mitigate CVE-2025-65887, organizations should first monitor for updates or patches from the OneFlow development team and apply them promptly once available. In the absence of official patches, implement input validation controls to detect and block tensors containing zero values before they reach the flow.floor_divide() function. Employ runtime application monitoring to detect abnormal crashes or DoS symptoms related to floor division operations. Restrict access to OneFlow services to trusted users and networks to reduce exposure to crafted inputs. Consider sandboxing or isolating AI workloads to limit the impact of potential crashes. Additionally, incorporate fuzz testing and static code analysis in the development lifecycle to identify similar arithmetic vulnerabilities proactively. Document and train AI engineers on secure coding practices related to numerical operations. Finally, maintain robust backup and recovery procedures to minimize downtime in case of exploitation.