Stage – Security of autonomous AI coding agents: experimental evaluation of attacks and defences

Durée du stage : 6 mois
Date de démarrage envisagée : février 2027
Localisation du poste : Palaiseau (91)
Gratification : 1200 € net mensuel
Référence de l’offre à mentionner dans l’objet dans votre e-mail de candidature : STAGE-2027-08-Explo

Internship context

Join the Institut de Recherche Technologique SystemX, located at the heart of the world-class Paris-Saclay scientific campus, and take an active part in building an international technology research centre in digital systems engineering. Backed by France’s leading research organisations in the field and staffed by joint teams from industry and academia, the centre generates new knowledge and technological solutions based on breakthroughs in digital engineering and disseminates its expertise across all economic sectors.

You will be co-supervised by two SystemX research engineers in cybersecurity and artificial intelligence, and by a researcher from Argonne National Laboratory (USA).

You will work within a SystemX exploratory research project on the security of AI agents, in academic partnership with Argonne National Laboratory (USA).

The position is based at IRT SystemX – Palaiseau.

Detailed description

AI coding agents are increasingly used to write, test and run code autonomously. To do so, they execute commands and modify files on behalf of developers, often with broad privileges, while processing content from sources they do not control: documentation, third-party dependencies, issues or web pages. This makes them an attractive target: malicious instructions hidden in that content (indirect prompt injection) can divert an agent from its task, with potentially serious consequences for the code, the data and the systems it has access to. The threat is now documented by real-world incidents and vulnerabilities, and securing these agents has become a prerequisite for their deployment. Many protection mechanisms have been proposed, both by industry and academia, but their actual effectiveness against determined attackers, and their impact on the usefulness of the agents, remain insufficiently understood.

Internship objectives

The internship aims to experimentally evaluate the security of autonomous coding agents: build reproducible attack scenarios, measure how current protection mechanisms withstand them and at what cost and explore possible improvements.

Tasks

• Review the state of the art on the security of LLM agents: threats, defence mechanisms and evaluation methods.
• Set up an isolated and instrumented experimental environment for a selection of coding agents.
• Design and implement reproducible attack scenarios and evaluate the agents and their protection mechanisms in terms of both security and utility.
• Explore and prototype defence mechanisms and assess their benefits and limitations.
• Analyse and document the results, contribute to a scientific publication, and present the work to the team and the academic partner.

References

• Greshake K., Abdelnabi S., Mishra S., Endres C., Holz T., Fritz M., “Not What You’ve Signed Up For: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection”, ACM AISec, 2023.
• Debenedetti E. et al., “AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents”, NeurIPS Datasets and Benchmarks Track, 2024.
• Nasr M., Carlini N. et al., “The Attacker Moves Second: Stronger Adaptive Attacks Bypass Defenses against LLM Jailbreaks and Prompt Injections”, arXiv:2510.09023, 2025.
• OWASP GenAI Security Project, “OWASP Top 10 for Agentic Applications”, 2026.

Profile and skills

Education: final year master’s or engineering school student (MSc level) in cybersecurity, computer systems or artificial intelligence.

Required skills :

• Python, shell scripting, Git
• Linux systems, containers and virtualisation
• Software and systems security: threat models, vulnerabilities, offensive testing in a controlled environment
• LLMs and agents (tool calling, MCP)
• Reproducible experimental methodology, analysis of measurements

Personal qualities :

• Rigour and autonomy
• Curiosity and interest in research
• Sense of ethics and confidentiality (responsible disclosure)
• Fluent English (co-supervision with a US partner)
• Good writing skills, interest in scientific publishing