Thesis topic:

Fixing Guesswork and Repetitive Findings in AI Security Interviews

Supervisor: Vjatšeslav Antipenko
- contact: vjatseslav.antipenko@ut.ee
The Problem: When conversational AI is used to interview business staff about their systems, two common issues arise:
- Guessing: If the user's answers are brief or vague, the AI starts to assume problems that were never mentioned (e.g., claiming that backups or two-factor authentication are missing without evidence).
- Duplicate Findings: The AI often breaks a single security flaw into multiple separate tickets (e.g., listing one weak password as a login flaw, a system takeover risk, and a data leak alert). This fills reports with repetitive noise that humans must manually clean up.
Research Goal & Approach: You will build and test two types of guardrails for our multi-agent assistant:
- An automated check during the interview that flags missing facts and stops the AI from guessing.
- A post-processing step that detects overlapping findings and groups them into clean, single summaries.
Deliverable: A functional software module that plugs into the agent pipeline to filter out guesses and merge duplicate findings, backed by a test demonstrating the reduction in noise.