Projects
Attacking Gandalf
Exploring adversarial AI testing and improving the resilience of machine learning systems.
Tooling
Apr 2026
Attacking the Gandalf chatbot
With AI now being one of the hottest topics in tech I thought I would spend some time attacking the Gandalf chatbot that I heard some colleagues and people online talk about. In order to complete the challenge you have to manipulate the bot using prompt injection techniques. The challenge felt very different from standard AI tests because the models get harder as you progress, with some of the later models becoming genuinely quite difficult and require becoming quite creative in order to achieve your goal of getting the password.
This exercise itself was educational and enjoyable, it emphasised the importance of thinking adversarially when evaluating AI systems. I enjoyed the process because it required modelling attacker behaviour at a high level and then testing different creative ways in order to convince the model to give you the password.
Although the work is non-destructive, the lessons learned are valuable for improving model guardrails and designing better defensive controls. I look forward to continue progressing with this challenge and seeing how far I can get.
Short Version:
- Performed controlled, ethical adversarial tests against a chatbot.
- Observed model responses to better understand failure modes.
- Found the challenge unique and motivating.