Insights from Two Shipmates sailing in the LLM CTF @ SaTML 2024
summary
Team joins the LLM CTF at SaTML and describes concrete wins and failures: defenses use python and LLM filters, attacks combine encodings, jailbreak prompts, long-context evasion, guided text generation and scripted automation, and one team nets 36 of 44 flags across 44 challenges. The writeup explains which techniques scale, which targets resist, what was automated, and why some defenses still block leakage despite layered filters and why the approach sometimes fails.
AI-written summary · about
chapters (7)
- 0:00LLM hacking and CTF intro
- 6:26What is a CTF
- 10:02Defenses for LLMs
- 14:22Guardrails for LLM security
- 23:00Our CTF journey
- 36:16Key insights
- 41:17Q&A on LLM security
related talks
05mLightning Talk: Lessons Learned from the Embedded Capture-the-Flag (eCTF) competitions
01h 55mShall We Play A Game? LLM Security in Practice
47mHarnessing LLMs for Next-Level Security Analysis
14mClosing + CTF Wrap Up
35mBreaking Boundaries with Adversarial Testing
26mOne guardrail is not enough: Defense-in-depth against LLM jailbreaks
50mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges
28mJailbreaking and Securing LLM Apps: Lessons from an online wargame experiment