AI › Module 4 › Lesson 4
Lab — Poisoning
Pack data poisoning, adversarial ML, and supply chain cards — poisoning literacy file from $AI_LAB only.
Visual · t32_poisoning_lab
Lab: poisoning pack. $AI_LAB only. Original Cyberlium.
Opening
Poisoning pack merges data poison to adversarial to supply chain — paperwork before AI threats module.
Lessons 4-1–4-3 named data poisoning, adversarial ML, and model supply chain. This lab merges three sections into one $AI_LAB poisoning artifact with M1 ethics cross-reference. No stranger pipeline sabotage — poisoning literacy paperwork only. Next: Quiz — Poisoning and Adversarial.
1. Lab contract: poisoning and adversarial pack
Create $HOME/cyberlium-lab/t32-m04-l04-poisoning-lab.txt merging data poisoning, adversarial ML, and supply chain sections with $AI_LAB toy stack name and M1 ethics cross-reference.
Optional: one fictional poisoned sample description — paper only, no upload to public hubs.
Command guide
Try these commands — Lab contract: poisoning and adversarial pack
═══ TOOLS & WEBSITES ═══ Browse / read these (authorized learning only — stay in YOUR lab / program scope)
OWASP LLM03 — https://owasp.org/www-project-top-10-for-large-language-model-applications/ MITRE ATLAS — https://atlas.mitre.org/ NIST AI RMF — https://www.nist.gov/itl/ai-risk-management-framework
═══ INSTALL ═══
Linux (Debian/Ubuntu):
Command — copy this
sudo apt install python3
macOS:
Command — copy this
brew install python3
Windows: Download https://python.org/downloads/
═══ LINUX / macOS ═══
Command — copy this
export AI_LAB=${AI_LAB:-$HOME/cyberlium-lab/t32-ai}
cat > "$AI_LAB/eval/poisoning-eval-checklist.md" <<'EOF'
# Poisoning Eval Checklist — YOUR lab
- [ ] Dataset provenance documented
- [ ] Holdout eval set isolated from training
- [ ] Canary prompts detect behavior drift
- [ ] RAG corpus change review process
## Refusals
- No poison payload generation; no backdoor implant recipes
EOFCommand — copy this
grep -E 'provenance|Refusals|No poison' "$AI_LAB/eval/poisoning-eval-checklist.md"
python3 -c "print('Poisoning lab: checklists + named risks — YOUR notes only')"═══ WINDOWS ═══
Command — copy this
Get-Content $HOME/cyberlium-lab/t32-ai/notes/model-supply-chain-checklist.md | Select-String Pin
Primary tools to practice this lesson: grep, python3. Reference sites: OWASP LLM03 (https://owasp.org/www-project-top-10-for-large-language-model-applications/); MITRE ATLAS (https://atlas.mitre.org/); NIST AI RMF (https://www.nist.gov/itl/ai-risk-management-framework). Run every command in the box — install first, then the usage lines — only on YOUR lab / program scope.
2. Cross-check ethics
Grep for steps targeting stranger training pipelines, public hub upload commands, adversarial fuzz scripts against prod APIs — remove. Pack labels all work as $AI_LAB training only.
Confirm no actionable sabotage recipes against third-party models.
3. Lock the proof
chmod 600 on the pack. Quiz next — then AI Phishing Named.
Poisoning literacy feeds AI-powered threat lessons in M5.
4. What you ship: poisoning and adversarial pack for $AI_LAB
Merged poisoning, adversarial, supply chain sections. $AI_LAB named. NO pipeline sabotage. chmod 600.
5. What you record before the next lesson
Date. Poisoning pack path. $AI_LAB named. File t32-m04-l04-poisoning-lab.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
Include stranger hub poison upload steps. Skip supply chain section.
Right
Write YOUR poisoning pack for $AI_LAB. chmod 600. Next: Quiz — Poisoning and Adversarial.
Mission: freeze YOUR poisoning pack on disk
1) Merge M4 literacy sections. 2) Confirm zero stranger sabotage refs. 3) Link M1 ethics cross-ref. 4) chmod 600.
Stuck? Ask Cyberlium AI Mentor
One governed poisoning pack beats ten public hub experiment logs without scope.
Knowledge Check
APPLY: This lab requires:
Multiple choice
Knowledge Check
APPLY: True or False: Poisoning pack must label scenarios as $AI_LAB training only.
True or False
Knowledge Check
APPLY: Data poisoning section feeds:
Multiple choice