Project · Threat model + risk register for one real LLM feature

Pick one real or realistic LLM feature (a chatbot, RAG assistant, or tool-using agent). Draw its attack-surface map, run a STRIDE-for-LLM pass producing at least six abuse cases, tag each with an OWASP LLM Top 10 category, and load them into a scored risk register ranked by priority.

• An attack-surface map listing every trust boundary the feature has, with who controls each one and whether it's trusted by default.
• A STRIDE-for-LLM table with at least six concrete, feature-specific abuse cases.
• A risk register (code or spreadsheet) with each risk scored likelihood x impact and tagged to an OWASP LLM Top 10 category.
• A one-paragraph writeup naming the single highest-priority risk and what you'd do first to reduce it.

• The attack-surface map lists at least four distinct surfaces, not just the chat input box.
• Every abuse case is concrete and specific to the chosen feature, not a generic restatement of the OWASP category name.
• Every risk in the register has a defensible likelihood and impact score and maps to a real OWASP LLM Top 10 category.
• The writeup's top-priority risk is genuinely the highest-scored item in the register, and the proposed first action is specific.