Red Teaming of LLM and Agent-Based Applications

Fixed-term contract (typically 3 months) · 80–100% · Zürich · Students (Bachelor, Master or PhD level)

«Making sense of data is the winning competence in all industries.»

We build D ONE with the belief that every company will be a data company. Since our beginnings, we help our clients to make sense of their data and generate value.

In this internship you will work independently on a topic in the area of AI security and safety testing. You will probe generative AI systems for weaknesses, and turn what you learn into tooling and guidance that others can reuse.

Your tasks will include:

• Reviewing the current landscape of attacks and defences for LLM and agent-based applications.
• Designing and implementing adversarial test cases and automating their execution in Python.
• Assessing systems against recognised security and governance frameworks.
• Analysing findings, judging their severity and proposing mitigations.
• Documenting your work and presenting it to a technical and a non-technical audience.

What you bring:

• Current enrolment in a Bachelor's, Master's or PhD programme in Computer Science, Data Science, Computational Linguistics, Engineering, Mathematics or a comparable field.
• Interest in at least one industry domain and curiosity about how data creates value in it.
• Down-to-earth and pragmatic results-oriented attitude.
• Experience in working with large data sets and databases.
• Knowledge in relevant technologies and programming languages (e.g. Python, SQL, Git, Spark, cloud platforms).
• Excellent communication skills in English; German is a strong plus.

Preferred qualifications:

• Working understanding of large language models, retrieval-augmented generation and tool-using agents.
• Familiarity with LLM-specific attack classes: prompt injection (direct and indirect), jailbreaks, data and system-prompt exfiltration, insecure output handling, excessive agency and tool misuse.
• Awareness of the OWASP Top 10 for LLM Applications, the NIST AI Risk Management Framework and the EU AI Act.
• Exposure to red teaming or security testing tooling such as Garak, PyRIT, promptfoo or equivalent, or general penetration-testing experience.
• Ability to automate attack generation and scoring, including adaptive and multi-turn strategies.
• Knowledge of defensive measures: input and output guardrails, sandboxing, least-privilege tool design, content filtering.
• Software engineering discipline: extensible Python frameworks, plugin architectures, automated test suites, CI.