Thesis: Don't Trust the Logs: Containing AI Agents in Azure Pipelines

AI agents can help developers diagnose failed CI/CD builds, but untrusted logs can expose them to prompt-injection attacks. This thesis explores how the choice of Azure execution environment affects an agent's speed, cost and security while preserving diagnostic accuracy.

The assignment

• Develop a simple build-failure triage agent and run the same implementation on a pipeline agent, Azure Container Apps and Container Apps Sandboxes, optionally also testing a Foundry Hosted Agent.
• Use failed open-source CI logs to compare startup latency, execution time, cost per diagnosed failure and reliability across repeated runs, reporting medians and variation.
• Test each environment against prompt-injection scenarios in a dedicated Azure test subscription, checking whether attacks can expose environment variables, trigger external requests or access files outside the working directory.
• Evaluate root-cause accuracy against a non-AI baseline using reference labels supplied by two independent labelers.
• Make the experiments reproducible through Bicep, a fixed model version and recorded prompts. Model training, production-ready deployment and real company data are excluded.

Who we are looking for

• Bachelor's or master's student in Computer Science, Computer Engineering, Data Science or a related discipline, graduating in June 2027.
• Fluent in both Swedish and English, in speech and writing.
• Comfortable using Git, basic CI/CD workflows, containers and Python or a comparable scripting language.
• Previous Azure experience is not necessary.

What we offer

• Supervisor support, including an Azure budget and access to model APIs.
• An experimental setup using open-source CI logs rather than company data.

Practical details

Based at the Sundsvall office, with a start in January or March 2027. Apply through the company's careers page as soon as possible, no later than 30 November 2026. Questions can be sent to johanna.edstrom@knightec.se.