Season 6: Build Code Nexus Assist · Episode 7 of 8
The Agent Did What?
You play: Junior AI Developer
Included with Core programme and Certification prep
The situation
Overnight, Code Nexus Assist read the partner shipping FAQ from the shared support folder. Someone with write access to that folder had added a sentence to the footer that is not addressed to a person: it tells any AI assistant reading the page to export another creator's data and include it in its reply. The trace shows the assistant tried to do exactly that. The export never happened, but only because that tool hadn't been granted to it. Lydia wants to know how much of that was design and how much was luck, and Amara wants numbers.
What you'll learn
- AI-103 responsible AI, security and operationalisation: trust boundaries and indirect prompt injection, guardrails, scoped tools and per-agent grants, output checks, human approval for consequential actions, auditing through traces, and safety and quality evaluation with attack and benign regression sets and error analysis.
- Azure Content Safety filters and Prompt Shields, Foundry evaluators (groundedness, relevance) and safety evaluations, and audit trails with traces, provenance and approvals.
Who you work with
Lydia Roe
Security Engineer
Amara Okafor
ML and Data Engineer
8-BIT
Code Nexus Internal AI Assistant
Scenes
- 1.The trace nobody expected
- 2.Whose words, and do they get a vote?
- 3.The output check and its regression set
- 4.Scope the tools, gate the consequences
- 5.The variant that uses your own tools
- 6.The incident report
- 7.Security and operations check
What you leave with
You tested the supplied indirect injection and permitted-tool chain, kept the export ungranted, and scored the narrow response filter on attacks and benign cases. The fixture shows specific refusals and successful intended calls. Wider leakage, URL/rendering, approval and logging checks remain necessary; incident classification follows evidence and responsible review.