1. Abstract
AI agents can help developers investigate bugs, analyse data and automate maintenance. Their access to databases, files and logs needs clear technical boundaries.

This practical workshop explores how to build and test those boundaries. Participants will apply database access controls, create pseudonymised database copies while preserving relationships, detect unreviewed schema changes, and remove sensitive information from logs before normal ingestion.

Exercises use PostgreSQL, Microsoft Presidio, Fluent Bit, Loki and Grafana, supported by an AI coding assistant. Participants will observe the original exposure, implement a change and run checks to verify its effect. Throughout the workshop, we will distinguish between masking, pseudonymisation, anonymisation and access control - and discuss the limitations of each.

2 Agenda
09:00 - 09:20 Environment checks, introduction and baseline demonstration
09:20 - 10:50 Securing database access: permissions, row-level security, API boundaries, aggregates and identity
10:50 - 11:00 Break
11:00 - 12:10 Database pseudonymisation: consistent tokens, preserved relationships and schema-change detection
12:10 - 13:00 Presidio: detecting sensitive data, inspecting missed entities and adding custom recognisers
13:00 - 14:00 Lunch
14:00 - 15:20 Log processing with Fluent Bit: privacy filters, policy changes and verification in Loki/Grafana
15:20–15:30 Break
15:30 - 16:00 Re-identification risks, architecture discussion, optional extensions and questions

Selected security layers will be implemented hands-on; others will be demonstrated. Kubernetes and the model broker are optional extensions.

3 Objectives
By the end of the workshop, participants will be able to:

• Distinguish AI instructions from access controls enforced by the application, database, and infrastructure.
• Apply and test selected PostgreSQL permissions, row-level policies, and API access boundaries
• Explain how process, filesystem, host and network restrictions limit an agent’s access.
• Generate consistent pseudonyms while preserving foreign-key relationships.
• Require an explicit processing decision for new database columns before publishing a refreshed copy.
• Configure and evaluate Presidio recognisers using annotated examples.
• Apply log privacy filters before ingestion and verify the stored results.
• Explain what passing tests demonstrate and identify remaining exposure or re-identification risks

4. Target audience and Prerequisites
The workshop is intended for:

• Backend and full-stack developers.
• DevOps and platform engineers.
• Software and solution architects.
• Security engineers working with application data or AI-enabled development.
• Software team managers.
  Participants should be comfortable running terminal commands, reading application code and using basic SQL, including SELECT and JOIN. Familiarity with Docker and an AI coding assistant will be helpful.

Exercises use synthetic data in local Docker environments. Kubernetes is discussed as an optional extension.

5 Technical requirements Participants should bring a laptop with the following installed and working

VS Code or another editor
Codex or Claude Code
Docker Desktop
DBeaver Community - PostgreSQL JDBC driver downloaded beforehand
Git
Python 3.10 or newer
Web browser
OpenAI API Key
Pre-pulled Docker images:
docker pull postgres:16
docker pull python:3.13-slim-bookworm
docker pull node:22-bookworm-slim