FDE Times: Red Teaming LLM Applications — keep the attacker’s mindset, drop the Giskard v2 notebooks
A 6 November FDE Times review of DeepLearning.AI’s ~1h39m Beginner course from Giskard staff says the attacker mindset and business-risk framing still hold; Giskard v3 is a rewrite, so course notebooks will not run as-is on a client project.
FDE Times reviewed DeepLearning.AI’s Beginner short course Red Teaming LLM Applications (seven videos, one graded assignment, ~1 hour 39 minutes), taught by Giskard’s Matteo Dora and Luca Martial and launched in early April 2024. The course centres on prompt-injection attacks against chatbots and on automating those attacks—skills every FDE needs when a client asks what happens if someone deliberately tries to break the demo.
Three ideas still travel well: treat red teaming as a business risk (legal and reputation), think like an attacker (aligned with OWASP LLM01), and automate at scale without trusting a single clean run—Giskard’s own docs warn a clean scan only means those specific attacks failed, and the grading LLM can err both ways, so failures need human review.
What to drop is the sample code: Giskard v2 is no longer actively maintained, and v3 is a ground-up rewrite (scan split into vulnerability_scan and quality_scan; async/Pydantic entry points). Watch the course for the mindset, then reimplement against current v3 docs and save the scenario suite for CI when prompts or models change.
Source: fdetimes.net