Software Engineering Radio - the podcast for professional software developers SE Radio 692: Sourabh Satish on Prompt Injection
14 snips
Oct 28, 2025 Sourabh Satish, CTO and co-founder of Pangea and seasoned security entrepreneur, breaks down prompt injection risks for LLMs. He covers the OWASP Top 10 for generative AI, the $10K prompt-injection challenge and attacker techniques like obfuscation and style injection. Learn about ingress/egress filters, system-prompt guardrails, detection methods, and priorities for securing LLM deployments.
AI Snips
Chapters
Transcript
Episode notes
Layered Defenses Dramatically Cut Success Rates
- Simple system-prompt guardrails are trivial to bypass: about 1 in 5 players succeeded.
- Stronger layered defenses (input/output inspection plus injection detection) reduced success to nearly zero.
Creative Token Tricks Evade Filters
- Attackers use distractor instructions, cognitive hacking, and style injection to evade detectors.
- Tiny token tricks (spaces, newlines, hyphens) and encoding change token interpretation and bypass filters.
Use Ingress And Egress Filters
- Apply ingress filters to inspect user input and contextual data before sending to the LLM.
- Apply egress filters to redact or block sensitive outputs like social security or full credit card numbers.