Playbook · for Engineering, Security
AI supply chain: models, packages and tools
Models, LLM frameworks and MCP servers are dependencies. Inventory them, pin them, scan them and keep their keys out of code.
By Mr.DLast reviewed 2 min read
What the AI supply chain is
Most of an AI feature is code and data you did not write: the model, its weights and tokenizer, an LLM framework, vector database clients, MCP servers, datasets, and the API keys that connect them. OWASP lists supply chain risks as LLM03. In practice, the most common entry points we see are vulnerable agent frameworks and MCP servers, and leaked model API keys.
1. Know what you run
Build an inventory, the same way you would for any dependency:
- models and their sources (vendor API, hosted, self-hosted weights);
- LLM frameworks, inference servers and their versions;
- MCP servers and plugins used by agents and coding assistants;
- datasets used for fine-tuning or retrieval, with their origin and licence;
- every API key and who owns it.
A software bill of materials (SBOM) from your build covers the code part. The EU Cyber Resilience Act will expect one for products sold in the EU.
2. Pin and verify
- Pin exact versions of LLM frameworks and MCP servers, and update deliberately.
- Download model weights only from trusted publishers, verify checksums or signatures, and prefer safe formats such as safetensors over formats that can execute code when loaded.
- Mirror critical packages and models in your own registry, so a deleted or hijacked upstream does not change what you deploy.
3. Scan continuously
- Scan dependencies for known vulnerabilities in every pull request, and again on a schedule, because new CVEs land on old versions.
- Watch advisories for the frameworks you use. Agent frameworks, inference servers and MCP servers have had a steady stream of critical flaws in 2026.
- Scan container images for the inference and agent runtimes too.
4. Keep keys out of code
Leaked model keys are resold for LLMjacking: attackers run inference on your account and you get the bill.
- Store keys in a secrets manager and inject them at run time.
- Scan commits and history for secrets in every pull request, and add a pre-commit hook.
- Rotate keys automatically, and immediately when one was ever committed, even briefly.
- Set spend limits and alerts on every model provider account.
5. Check your AI vendors
For each AI provider, record where data is processed, how long it is retained, whether it is used for training, and how the vendor will tell you about an incident. AI platform breaches in 2026 exposed customer datasets and secrets, so plan how you would rotate everything a vendor held.
6. Isolate what you cannot fully trust
Run third-party models and MCP servers with network isolation and minimal permissions, so a compromised component cannot reach the rest of your environment.
What WardBee checks
With code scanning turned on for a GitHub or GitLab repository, WardBee finds secrets in code and recent history, vulnerable dependencies and insecure infrastructure code. On AWS it checks for secrets in Lambda and CodeBuild environment variables and Secrets Manager rotation. Results map to ISO 27001, SOC 2 and the EU CRA.
What WardBee checks for this playbook
- Lambda function environment variables do not contain secrets
awslambda_function_no_secrets_in_variables - CodeBuild project has no sensitive credentials in plaintext environment variables
codebuild_project_no_secrets_in_variables - Secrets Manager secret has rotation enabled
secretsmanager_automatic_rotation_enabled
Runs on each AWS sync while a framework it maps to, such as ISO 27001, is enabled. Run it on your account
Related standards: OWASP Top 10 for LLM Applications · EU Cyber Resilience Act · ISO/IEC 27001


