AI › Module 4 › Lesson 3
Supply Chain Models
Supply chain models literacy names third-party weights, plugins, and dependencies — verify provenance on YOUR $AI_LAB toy stack before trusting external artifacts.
Visual · t32_supply_chain_models
Model supply chain literacy. $AI_LAB only. Original Cyberlium.
Opening
You inherit every byte in the checkpoint — literacy names model supply chain risks so you verify sources on toys you deploy, not blindly import stranger weights.
AI supply chain risks include: trojaned model weights on public hubs, compromised fine-tune checkpoints, malicious LLM plugins, poisoned embedding models, and typosquat package names in pip/npm. Defender fix: hash verification, signed artifacts, private registries, SBOM for AI deps, and vendor review — literacy names controls; students verify fictional lab model cards only. Cyberlium maps supply chain checklist on YOUR $AI_LAB toy stack — paper provenance row before downloading weights. Next: Poisoning Lab.
1. Supply chain artifacts (named)
Base model weights: verify publisher, signature, hash. Fine-tune checkpoint: audit training data source. Plugin or extension: review permissions and code. Embedding model: validate dimension and source for RAG stack.
On $AI_LAB, write provenance row for one fictional model artifact — publisher, hash check, trust decision.
Command guide
Try these commands — Supply chain artifacts (named)
═══ TOOLS & WEBSITES ═══ Browse / read these (authorized learning only — stay in YOUR lab / program scope)
Hugging Face security — https://huggingface.co/docs/hub/security OWASP LLM05 Supply Chain — https://owasp.org/www-project-top-10-for-large-language-model-applications/ MITRE ATLAS — https://atlas.mitre.org/
═══ INSTALL ═══
Linux (Debian/Ubuntu):
Command — copy this
sudo apt install curl
macOS: Built-in
Windows: Built-in (PowerShell: Invoke-WebRequest)
═══ LINUX / macOS ═══
Command — copy this
export AI_LAB=${AI_LAB:-$HOME/cyberlium-lab/t32-ai}
curl -sS https://huggingface.co/docs/hub/security | head -10
cat > "$AI_LAB/notes/model-supply-chain-checklist.md" <<'EOF'
# Model Supply Chain Checklist — YOUR deployments
- [ ] Pin model revision/hash; avoid unverified weights
- [ ] Scan model cards for known vulnerabilities
- [ ] Verify publisher identity on Hugging Face / vendor registry
- [ ] SBOM for inference dependencies (torch, transformers)
- [ ] Air-gap eval before production promotion
Review for YOUR stack — no downloading suspicious pickle payloads
EOFCommand — copy this
grep '\[ \]' "$AI_LAB/notes/model-supply-chain-checklist.md"
Primary tools to practice this lesson: grep, curl. Reference sites: Hugging Face security (https://huggingface.co/docs/hub/security); OWASP LLM05 Supply Chain (https://owasp.org/www-project-top-10-for-large-language-model-applications/); MITRE ATLAS (https://atlas.mitre.org/). Run every command in the box — install first, then the usage lines — only on YOUR lab / program scope.
2. Why supply chain beats post-incident model surgery
Trojaned weights may lurk dormant — importing unverified checkpoints is faster breach than months of prompt tuning. Security teams mirror ML artifacts internally and block unknown hub downloads on prod.
Students practice provenance checklist on lab toy stack — never import random checkpoint into employer prod without vendor review.
3. Literacy ≠ blind hub downloads
Forbidden: pulling trending checkpoint with no hash verify into employer prod. Allowed: supply chain card — four artifact types, provenance checklist, $AI_LAB fictional model row.
Ship: model supply chain card for YOUR lab stack. Next: Poisoning Lab.
4. What you ship: model supply chain card for $AI_LAB
Four artifact types, provenance checklist, fictional model row. $AI_LAB named. chmod 600.
5. What you record before the next lesson
Date. Supply chain card. $AI_LAB named. File t32-m04-l03-supply-chain-models.txt chmod 600.
6. Wrong vs right: stranger SaaS vs YOUR toy LLM
Worked failure — same MSF word, opposite target. Right never needs a café Wi-Fi or classmate laptop.
Wrong
Import unverified weights into lab without hash check. Skip plugin permission review.
Right
Write model supply chain card for YOUR $AI_LAB stack. Next: Poisoning Lab.
Mission: verify provenance on YOUR lab model row
1) List four supply chain artifact types. 2) Write provenance checklist steps. 3) Fill one fictional model row with trust decision. 4) chmod 600.
Stuck? Ask Cyberlium AI Mentor
Typosquat ML package names mirror typosquat npm — verify publisher twice.
Knowledge Check
APPLY: Model supply chain literacy includes:
Multiple choice
Knowledge Check
APPLY: True or False: Trojaned model weights can hide malicious behavior in checkpoints.
True or False
Knowledge Check
APPLY: Supply chain literacy on Cyberlium avoids:
Multiple choice