Protect your generative AI applications against jailbreaks
summary
Short practical checklist for defending prompt injection: shows concrete attacks (affirmative instructions, base64 encoding, low-resource language bypass, punctuation/optimization tricks) and concrete defenses — Amazon Comprehend trust-and-safety classifier, guardrails with input/output filtering and PII redaction, and perplexity scoring with a threshold. Recommends prompt engineering (reminder prompts), limiting context windows, separating trusted vs untrusted inputs, and picks Anthropic-style constitutional models as harder to jailbreak. Useful if needing operational controls now.
AI-written summary · about
chapters (5)
- 0:00Introduction to jailbreaks
- 0:57How language models work
- 4:56Tactic and bypass techniques
- 8:17Reasons for jailbreak success
- 11:46Defensive strategies
related talks
26mOne guardrail is not enough: Defense-in-depth against LLM jailbreaks
51mPrompt Injection: When Hackers Befriend Your AI
53mHow to Break AI Systems (Before Someone Else Does)
01h 01mHow to Break AI Systems (Before Someone Else Does)
29mSurvival of the Safest: Automating LLM Defense with Genetic Programming
30mBattling the Machine: Cutting-Edge Attacks and Defenses in Generative AI
37mKeynote - Hacking AI How to Survive the AI Uprising
18mDefending the unknown using AI and cybersecurity