APPENDIX H: RESUME PORTFOLIO BUILDING CAPSTONE PROJECTS

ARTIFICIAL INTELLIGENCE ASSISTANT PROMPTS These two prompts reflect security work a harness engineer does when connecting outside tools. Replace each bracketed placeholder with your own material before use, and share only descriptions, never secrets. [START OF PROMPT] I am connecting an agent to an outside tool server. Here is a plain description of the tools it offers, what each one can read or change, and what the agent is currently allowed to do: [PASTE TOOL NAMES, DESCRIPTIONS, AND CURRENT PERMISSIONS WITHOUT ANY SECRETS]. Act as a security reviewer and produce a short threat model. For each tool, list what could go wrong if tool output contained hidden instructions, if the tool description were altered after I approved it, or if the arguments were malicious. Then recommend which tools should be on an allowlist, which need human approval, and which should be refused. Output format: a list of tools, each with Risks, Recommended Policy, and One Test I Should Write. [END OF PROMPT] [START OF PROMPT] Here are the phrases my agent currently scans for to detect prompt injection in tool results: [PASTE YOUR PHRASE LIST]. Critique this approach honestly. Explain why phrase lists alone are weak, give me ten harmless example sentences, which only describe the shape of an attack without being real exploits, that a naive list would miss, and suggest three additional layers that do not depend on guessing phrases. Output format: Weaknesses, Ten Harmless Examples, Three Extra Layers. [END OF PROMPT] PROJECT 4: THE HARNESS REPORT CARD (AN EVALUATION AND OBSERVABILITY SUITE)