Pentesting AI: How to Hunt a Robot
summary
Six tangible tests target AI leaks, losses and modifications: try negative guidance and membership inference to pull private records from models, place bollards or projections to cause availability failures in driverless fleets, and use prompt injection or tiny token obfuscations (ch1ldren → ch1ldren) to bypass filters. Real incidents include Cruise shutdowns, an Uber fatality and Redis cache leaks that exposed chats; mitigation centers on personal data stores and hardened architectures, not blind trust.
AI-written summary · about
chapters (57)
- 0:00Introduction to robot hunting
- 0:52Bias in machines
- 1:34The impact of systemic bias
- 2:53Terminology in security
- 3:56Defining AI and cyber
- 7:14Intelligence and xenophobia
- 7:57The history of eugenics
- 8:36Artificial bias defined
- 9:01Testing methodologies
- 9:40Rebranding CIA triad
- 10:27Leaks and guidance
- 11:08Losses and swarms
- 11:48Prediction inversion
- 12:56Case study: Cruise vehicles
- 13:40Algorithmic risks
- 14:08Resource exhaustion
- 14:54The visual gap
- 15:39AI object recognition
- 15:57Fencing and breaches
- 16:13Lack of privacy visibility
- 16:41The tortoise test
- 17:02Modifications and prompts
- 17:29Dan and jailbreaking
- 17:39Automated hate and bias
- 19:21Input manipulation
- 19:39Contextual understanding
- 20:04Algorithmic murder
- 21:19Uber and pedestrian risk
- 21:58Palantir and accuracy
- 22:38Law enforcement views
- 24:02The threat of AI
- 26:40Learning to be bad
- 27:14Diminishing freedom
- 27:33Ingredients of genocide
- 27:45Institutional trust
- 28:15The utopia trap
- 28:32Information warfare history
- 30:13The Titanic lesson
- 34:30Weisenbaum's warning
- 35:00Intelligence vs understanding
- 36:20Replica and abuse
- 37:46Shifting left in security
- 38:41Full imagination testing
- 39:272007 gun failure
- 40:10Pen testing today
- 42:03Integrity failures
- 43:02Redis and architecture
- 44:37Banning and books
- 50:29Personal data stores
- 51:01Optimistic architecture
- 52:36The firewalls of the future
- 53:06Resetting the system
- 53:46Increased pen test impact
- 54:26Frankenstein and hype
- 55:42Education for understanding
- 57:38The need for ethics
- 58:05Conclusion
related talks
24mHow to Build Trustworthy AI
57mAI Applications Security Puzzle
45mSecuring AI: There Is No Try, Only Do!
53mHow to Break AI Systems (Before Someone Else Does)
35mHacking AI Systems: How to (Still) Trick Artificial Intelligence
18mDefending the unknown using AI and cybersecurity
54mA Practical Guide to AI Threats & Mitigation Strategies
44mHacking AI - How to Survive the AI Uprising