Executive Summary
The study found widespread security issues reported across AI project ecosystems, including dependency, model, data, and component-isolation problems. AI developers, platform owners, and security teams should prioritize supply-chain visibility and control validation.
What Happened
A study analyzed 312,868 security discussions and identified 32 issue types across system software, tools and ecosystems, models, and data. Frequently reported risks included dependency incompatibilities, unsafe deserialization, malicious output generation, sensitive-data leakage, and weak component isolation.
Vulnerability / Attack Technique
AI supply-chain compromise
Technical Mechanism
The reported issues span dependencies, tools, ecosystems, models, and data used by AI projects. Unsafe deserialization, malicious output generation, sensitive-data leakage, and weak component isolation represent distinct failure modes that can affect how AI components are loaded, executed, or connected.
AI Security Category
LLM03:2025 Supply Chain
Part of the AI Supply Chain and AI SBOM topic.
Relevant Frameworks
RealGround Analysis
AI agents and LLM applications often depend on interconnected software, model, data, and tooling components, so weaknesses outside the model can still affect system security. Teams should maintain an inventory of these dependencies and assess isolation, serialization, output-handling, and data-protection controls throughout the AI lifecycle.
Recommended Controls
- Maintain an AI-focused software bill of materials covering libraries, models, datasets, tools, and transitive dependencies.
- Scan and validate serialized model and software artifacts before loading them into development or production environments.
- Enforce least-privilege execution and strong isolation between AI components, tools, workloads, and sensitive data.
- Apply output validation and content-handling controls to detect and contain malicious model-generated output.
- Monitor AI project dependencies and component changes, with documented remediation and incident-response procedures.
Related RealGround Service
AI Agent SBOM and LLM Mapping · or start with the free AI Agent Security Scanner.
Sources
ACM: https://dl.acm.org/doi/full/10.1145/3744916.3787796
Summary and analysis written by RealGround from the source above; last updated 2026-11-11.
