David Bombal #571: Google Big Sleep: The End of Human Hackers?
Mar 31, 2026
Stephen Sims, an offensive security researcher and SANS instructor, returns with sharp takes on AI in cybersecurity. He unpacks offensive vs adversarial AI, prompt injection and jailbreak techniques. Vector databases, agentic automated testing, Google Project Zero’s Big Sleep, and AI-driven patch diffing get clear, bite-sized treatment. Practical career and governance implications wrap up the conversation.
AI Snips
Chapters
Transcript
Episode notes
Practice AI Attacks Only In Sanctioned CTFs
- Practice adversarial techniques like prompt injection and model extraction only in sanctioned environments.
- Stephen Sims points to CTFs and platforms (LaCara, Dreadnode Crucible) as safe places to test jailbreaks and injection attacks.
Agent Solved Gandalf CTF In Under Two Minutes
- Stephen Sims demos an agent solving the Gandalf CTF in under two minutes, showing agentic speed on certain tasks.
- The agent completed all levels quickly, illustrating how autonomous agents can outperform manual steps on scripted challenges.
Adversarial AI Targets Models Themselves
- Adversarial attacks include model extraction, prompt injection, poisoning, and LLM DoS targeting costs and logic.
- Sims explains model extraction queries and SpongeAttacks-style inputs that exhaust token/GPU resources as realistic threats.