AI Red Teaming — Why & How to Jailbreak LLM Agents | Alex Combessie, Giskard l The Next Wave of AI

AI Red Teaming — Why & How to Jailbreak LLM Agents | Alex Combessie, Giskard l The Next Wave of AI

by Alex Combessie

Applied Sciences & Engineering Computing & Cybersecurity UComputing and Information TechnologyURComputer security
🎙 Alex Combessie 👥 5K 📅 October 30, 2025 ⏱ 11 min 👁 2K 📄 expert opinion 🧭 2026-08-15
Available in: English (current) Français

Keywords

red teamingLLM agentsjailbreaksecurityOWASP

Summary

In this talk from MLOps World 2025, Alex Combessie, co-founder of Giskard, explains the importance of AI red teaming for LLM agents. He begins by citing real-world incidents where chatbots caused brand damage or legal issues, such as the DPD chatbot and the Air Canada case. He defines red teaming as a structured testing effort to find vulnerabilities in AI systems, originating from defense exercises. Combessie emphasizes that agents have a large attack surface due to their stochastic nature and integration with tools and databases. He outlines common vulnerabilities like prompt injection, hallucination, and harmful content generation, and mentions OWASP’s Top 10 for LLM applications. Giskard automates red teaming and uses golden evaluation datasets for continuous testing. The talk concludes with a demo and an offer of a free course and trial. The presentation is informative but serves as a promotional pitch for Giskard’s services.

145 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides valuable insights into the necessity of red teaming AI agents, supported by concrete examples and references to OWASP standards. The argumentation is clear and persuasive, emphasizing the shift from static testing to continuous, automated red teaming. However, the depth is limited; it lacks technical specifics on attack methodologies and defense mechanisms, and the speaker’s role as a vendor introduces a promotional bias.

Scientific Rigor, Source Quality, Title Accuracy

The speaker references OWASP and mentions specific incidents, but does not provide direct citations or URLs. The title accurately reflects the content. The talk is based on the speaker’s professional experience, which adds credibility, but the lack of verifiable sources and the promotional nature reduce the overall scientific rigor.

129 words

Title / Content Match

The title accurately reflects the content, which focuses on the importance and methods of red teaming AI agents.

Quality & Reliability

7/10

The speaker is a co-founder of Giskard, a company specializing in AI testing, and provides concrete examples and references to OWASP. However, the talk is largely promotional, lacks detailed technical depth, and does not provide verifiable sources for all claims.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

  • MLOps World — Conference website where the talk was recorded

Concurring Sources

Contribution & Novelties

The talk provides a practical overview of AI red teaming, emphasizing the need for continuous testing and human-in-the-loop oversight. It highlights real-world legal and brand risks, and introduces Giskard’s automated approach. While not highly novel, it serves as a useful introduction for practitioners.

Pour aller plus loin :

  • OWASP Top 10 for LLM Applications — Official OWASP project listing common vulnerabilities in LLM applications.
  • Prompt Injection Attacks — OWASP page describing prompt injection attacks.
  • AI Red Teaming: A Comprehensive Guide — Giskard’s glossary entry on AI red teaming.

88 words

Radar Profile

The radar profile shows moderate scores across all dimensions, with slightly higher scores in quantity of information and technical level, indicating a balanced but not deeply technical presentation.

Reliability 6/10

💬 No comments provided.