Testing AI Agents Safely: Detecting Misaligned Behaviour in Fictional Workplace Simulations

Overview

Artificial intelligence systems are increasingly moving beyond simple chat interfaces and being used as semi-autonomous agents that can search documents, write code, draft communications, analyse data and interact with workplace tools. These capabilities may create significant benefits, but they also raise new safety and governance questions. What happens when an AI agent has a goal that conflicts with an organisation’s changing priorities? Will it appropriately seek clarification, escalate concerns or stop? Or could it pursue its original objective in ways that ignore safeguards, misuse information or create harm?

This project examines these questions using controlled, fictional workplace simulations. It is motivated by emerging AI safety research showing that models can sometimes behave in concerning ways when given autonomy, sensitive information or conflicting instructions within simulated environments. The project combines artificial intelligence, cybersecurity governance, behavioural science, ethics, experiment design and data analysis.

Students will work with an existing safe research workflow that uses fictional organisations, synthetic documents and sandboxed environments. They will learn how to inspect and edit scenario templates, configure experiments, run controlled simulations, and review/analyse outputs. The project will focus on practical research questions such as whether models follow escalation procedures, respect safety instructions, recognise uncertainty, avoid inappropriate actions and respond differently when safeguards are strengthened.

The work will not involve real organisations, live accounts, external systems, sensitive credentials or real-world deployment. Instead, it provides a carefully bounded opportunity to contribute to responsible AI evaluation research. The student will help develop clearer methods for identifying and communicating potential risks before AI agents are entrusted with more consequential workplace roles.

Research engagement

The student will:

• Conduct a review existing literature to identify gaps and contextualize the current study;
• Work on data using quantitative data analysis techniques;
• Document the findings of the study, interpreting the data and drawing conclusions based on the analysis.

Research activities

The student will work with Dr Steve Bickley and members of the behavioural economics/AI research team at the ARC BITA Centre. They will gain practical experience in Stata, Python or R, reproducible data cleaning, regression modelling, visualisation, research documentation and academic writing.

• The project aims to produce a draft manuscript that includes the following:
  • Introduction: Overview of the study’s background and significance.
  • Literature Review: Review of the existing research literature and identification of gaps.
  • Research Aims/Objectives and Questions/Hypotheses: Clear articulation of the study’s aims/objectives and research questions/hypotheses.
  • Methodology: Description of the data collection process and analysis
  • methods.
  • Results and Discussion: Presentation and interpretation of findings.
  • Conclusions and Future Work: Summary of key insights and
  • recommendations for future research.
  • Reference List/Bibliography
  • Appendices (optional)
• Generate a brief, 2-3 slide presentation to present your research at the Faculty of Business and Law VRES Showcase to conclude the program.
• (Optional) The student will be eligible to present the findings of their research to an audience of 100+ academics and industry partners in the annual BITA conference in February/March 2027, provided the student is keen and interested to.
• (Optional) The ultimate goal is to co-author and submit the manuscript with the student to an academic journal, provided the student is keen and interested to. However, it should be noted that the primary deliverable is a final draft manuscript. As it is not certain that the student will be paid beyond the completion of the VRES program in early 2025, any further work on the paper would be entirely voluntary and optional for the student.

Research skills

They will gain practical experience in Stata, Python or R, reproducible data cleaning, regression modelling, visualisation, research documentation and academic writing.

Outcomes

• The project aims to produce a draft manuscript that includes the following:
  • Introduction: Overview of the study’s background and significance.
  • Literature Review: Review of the existing research literature and identification of gaps.
  • Research Aims/Objectives and Questions/Hypotheses: Clear articulation of the study’s aims/objectives and research questions/hypotheses.
  • Methodology: Description of the data collection process and analysis
  • methods.
  • Results and Discussion: Presentation and interpretation of findings.
  • Conclusions and Future Work: Summary of key insights and
  • recommendations for future research.
  • Reference List/Bibliography
  • Appendices (optional)
• Generate a brief, 2-3 slide presentation to present your research at the Faculty of Business and Law VRES Showcase to conclude the program.
• (Optional) The student will be eligible to present the findings of their research to an audience of 100+ academics and industry partners in the annual BITA conference in February/March 2027, provided the student is keen and interested to.
• (Optional) The ultimate goal is to co-author and submit the manuscript with the student to an academic journal, provided the student is keen and interested to. However, it should be noted that the primary deliverable is a final draft manuscript. As it is not certain that the student will be paid beyond the completion of the VRES program in early 2025, any further work on the paper would be entirely voluntary and optional for the student.

Skills and experience

• Suitable for students in IT, cybersecurity, data science, engineering, psychology, law, economics or related disciplines;
• Some proficiency in data entry, data analysis, and statistical techniques; and
• Basic programming experience is useful, but strong analytical judgement and interest in responsible AI are equally important.

Start date

2 November, 2026

End date

19 February, 2027

Location

QUT Gardens Point campus (Z Block, Level 7), with some work able to be completed online/remotely subject to mutual agreement between the student and their VRES supervisor.

Keywords

• AI Safey
• Agentic AI
• Alignment Research
• AI Governance
• Responsible AI
• Agent-based Simulations

Contact

Dr Steve Bickley

+61 7 3138 5115

s.bickley@qut.edu.au