A new report out today from software supply chain company JFrog Ltd. warns that an expansion of artificial intelligence technology across the software supply chain has resulted in an alarming rise in security threats.

The finding comes from JFrog’s 2025 Software Supply Chain State of the Union, released to coincide with the KubeCon + CloudNativeCon Europe conferences. The report highlights emerging software security threats, evolving DevOps risks, best practices and increasingly serious security concerns in the AI era.

Key findings in the report include that a “quad-fecta” of security vulnerabilities is threatening the software supply chain. The top security factors affecting the integrity and safety of the software supply chain include Common Vulnerabilities and Exposures, malicious packages, secrets’ exposures, and misconfigurations and other human errors.

In an example in the report, the JFrog Security Research Team detected 25,229 exposed secrets or tokens in public registries, up 64% year-over-year, of which 27% were active. The increasingly sophisticated and intertwined fabric of software security threats make it difficult for organizations to maintain consistent software supply chain security.

AI and machine learning model proliferation and attacks were found to be growing. In 2024, there were more than 1 million new models and datasets added to Hugging Face, the largest repository of public machine learning models, with an accompanying 6.5-times increase in malicious models.

Though publicly uploaded models are increasingly presenting risks, organizations manually governing machine learning models were also found to be increasing risks. Some 94% of organizations create certified lists of approved models to govern how developers use machine learning artifacts, but 37% of companies still rely on manual efforts to curate and maintain that list, creating trepidation around the accuracy and consistency of model security.

Binary scanning — the process of analyzing compiled software, or binaries, for security vulnerabilities and malicious code that may not be detectable in the source code — was found to be lacking. Only 43% of information technology professionals said their organization applies security scans at both the code and binary levels, leaving many organizations vulnerable to security threats only detectable at the binary level. That’s down from 56% in 2023, indicating that despite growing risks, security basics such as binary scanning are either being overlooked or intentionally not applied.

Other findings in the report included persistent issues with open-source security. More than 70% of developers continuing to download packages directly from public registries, a risky practice that can expose entire organizations through a single compromised machine. Additionally, critical software vulnerabilities are on the rise, with more than 33,000 new CVEs disclosed in 2024, up 27% year-over-year.

The report also highlights concerns over CVE mis-scoring, revealing that only 12% of CVEs rated as “critical” were actually exploitable, raising doubts about current scoring methods. Lastly, the growing use of multiple security tools — 73% of professionals report using seven or more — may be contributing to increased complexity and risk, suggesting that a streamlined, more focused approach could offer better protection.

Image: SiliconANGLE/Reve