Shadow AI Creates Three Distinct Risks Across Data, Code and Software Supply Chains
Summary
Shadow AI creates three distinct risks: sensitive data sent to external services, unverified AI-generated code and dependencies, and models entering software builds. The article outlines governance, review, scanning and proxy controls for each.
Key points
- The article separates Shadow AI into three exposures: data sent to external services, AI-generated code and dependency choices, and models or machine-learning packages entering software builds.
- According to Verizon's 2026 DBIR, 45% of employees regularly used AI on corporate devices, up from 15% a year earlier; two-thirds of those users accessed services through non-corporate accounts.
- For data exposure, the article recommends sanctioned services, data-loss prevention and network controls. It says DoW Instruction 8430.01 bars putting non-public departmental information into GenAI services unless the service is both on departmental systems and approved.
- Treat AI-generated code as unverified input: review it, run software composition analysis on introduced dependencies, and record the models, versions and significant datasets used.
- Sonatype Research Labs audited 9,658 Hugging Face requests in the first half of 2026; nearly 7% were blocked or rejected. Among 674 blocked models, the article says 55% were stopped for model acceptability and 18% on security grounds.
- For model intake, the article recommends routing downloads through a governed proxy, blocking direct corporate-network downloads, and starting with audit mode before quarantining models.
- The article cites joint government guidance recommending that organizations require an AI bill of materials from vendors as a procurement condition.
Article Details
- Defense Focus
- Govern Shadow AI across data disclosure, AI-generated code and dependencies, and AI models or machine-learning packages entering software builds.
- Detection Methods
- Use data loss prevention and network controls to identify or limit data sent to external AI services.
- Run software composition analysis on dependencies introduced with AI-generated code.
- Route model downloads through a proxy repository and review request activity in audit mode before applying quarantine.
- At scan time, identify model derivatives and their foundation models; track model lineage, commit hash, and publishing date.
- Maintain an inventory and records of the models, versions, and significant datasets used to generate or test software.
- Data Sources
- Data loss prevention and network-control telemetry
- Proxy repository model-pull requests
- Software composition analysis results
- Model and software inventories, including lineage, commit hash, and publishing date
- Software evidence packages and software bills of materials
- Platforms
- Sonatype Firewall
- Sonatype Lifecycle
- AI Models Usage dashboard
- Defensive Actions
- Name specific sanctioned AI services and provide a usable approved platform; give staff practical guidance.
- Require human review of AI-generated code and assess introduced dependencies with software composition analysis.
- Route model pulls through a proxy repository, block direct downloads from the corporate network, and lock client configuration to prevent bypass.
- Begin model-boundary monitoring in audit mode, review what teams retrieve, and then apply quarantine where appropriate.
- Require records of models, versions, and significant datasets used in software development, and seek AI bills of materials from vendors during procurement.
- Maintain a complete, continuously updated inventory of software, including models.
Vendors
Products
Sonatype FirewallSonatype Firewall can then audit or quarantine models at that boundary before a developer gets them.Sonatype LifecycleAt scan time, Sonatype Lifecycle flags a model as a derivative and names the foundation model it came from, and the AI Models Usage dashboard shows which applications are using models, with lineage, commit hash and