Ai Risks + Viva Engage

September 18, 2025 – Lab Notes

Research & Exploration

• Looked into open-weight models and guardrails:
  • Guardrails are implemented after weights are released, not bundled with them.
  • Risks: poisoned fine-tuning can create backdoored models.
  • Explored unauthorized tool access concepts:
    • Exfiltration = stealing data out of a system.
    • Pivoting = using a compromised system to move deeper into the network.
    • Adversarial examples = inputs designed to trick models into bad outputs.
• Studied operational evasion in threat models:
  • Attackers shape their behavior to slip past detection systems.
  • Red team note: knowing defender culture shows what they neglect → attack surface.
• Explored ML/DL stack:
  • ML → DL → Transformers → LLMs progression.
  • PyTorch and TensorFlow: frameworks to build and train ML/DL models (both Python-based).
  • R language: stats-heavy, strong for data analysis.
• Learned about corpora: large structured sets of text used to train NLP/LLMs.
• Checked HuggingFace.co: hub for models, datasets, and tools for NLP/LLMs.
• Reviewed NLP basics: field of AI for processing and understanding human language.

---

Sys Admin

• Investigated Viva Engage Native Mode upgrade failure in Microsoft tenant.
  • Logged issue: bouncing back to login when accessing admin center.