Skip to content
AI.info

The Pulse

CleanStart report flags risks in AI-recommended software dependencies

Only 20% of tested AI-generated dependency recommendations were both valid and free of known vulnerabilities, according to a CleanStart report. The report also examines aging components and vulnerability findings in audited commercial software.

CleanStart report flags risks in AI-recommended software dependencies

Only 20% of tested AI-generated dependency recommendations were both valid and free of known vulnerabilities, according to a report from software supply-chain security provider CleanStart. The finding sits alongside evidence of aging components and growing numbers of vulnerabilities in audited commercial software.

Source prnewswire.com

CleanStart released its State of Software Dependency Trust 2026 report on November 17. The report combines published audits, surveys, registry telemetry, incident reports and policy sources rather than presenting all the findings as a single test.

AI recommendations under scrutiny

Dependencies are software components used by an application, including packages brought in through other components. AI-assisted development can introduce recommendations for those dependencies as developers build and maintain software.

The report also states that grounding AI models in live registry data reduced high- and critical-risk recommendations by roughly 70%. It examines how security tools can improve the quality of AI-generated dependency recommendations.

In a separate study of dependency upgrade recommendations, security company Sonatype found that even the best ungrounded models fabricated about 6% of recommendations. Sonatype tested several models against an approach using real-time software intelligence, examining whether suggested versions and upgrade paths held up in software ecosystems.

Older and indirect components

Audited applications averaged 1,180 open-source components, with 64% introduced indirectly through other dependencies, CleanStart reported. Across 947 audited commercial codebases, 17% of observed components were outside conventional package managers and 41% were at least ten versions behind.

The same sample contained components of which 68% were more than two years old, according to the report. Black Duck's audit data covered commercial and proprietary codebases examined for transactions, regulatory compliance and internal risk assessments between November 2024 and October 2025.

The average number of vulnerabilities per audited codebase rose 107%, from 280 to 581, CleanStart reported.

Fixes and verification

Separate registry telemetry from 2024 showed that approximately 95% of vulnerable component downloads examined had a fixed version available when they were downloaded, according to the report. A review of 47 leading non-npm open-source projects also found that a small number of developers accounted for most commits in 40% of those projects.

The report recommends beginning verification with widely reused or consequential images, libraries, runtimes and packages. It calls for minimum evidence covering their identity, origin, provenance, integrity, security, maintenance and policy, and for making that evidence reusable across applications and accessible through developer and AI workflows.

CleanStart says it applies this model through Clean Images and Clean Libraries, which provide verified, reusable software foundations. The company describes the approach as complementing continuing scanning, software bills of materials, remediation and runtime controls.

Sources

  1. PR Newswire
  2. Black Duck
  3. Sonatype

Related stories

Explore

More articles