Uploaded March 2026 | Updated September 2026, 2 weeks ago
As AI moves from augmentation to full automation, the stakes for privacy and security have never been higher. In this InfoQ video, privacy expert Katharine Jarmul breaks down why current AI "guardrails" are easier to bypass than you think and why senior engineers cannot rely on model providers to solve privacy for them.
Learn the architectural realities of overparameterization, why models "memorize" sensitive data, and how to build a culture of psychological safety to catch incidents before they scale.
⏱️ Video Timestamps (For Navigation)
0:00 – The Shift: From Augmentation to Automation
2:15 – The Problem with Privacy "Fearmongering"
4:30 – Myth #1: Why Guardrails (Software & Algorithmic) Fail
7:45 – Bypassing Filters: Variable Renaming & ArtPrompt Attacks
10:12 – Myth #2: Overparameterization & The Death of Overfitting
13:20 – Can LLMs Learn Without Memorizing? (The Data Leak Risk)
15:45 – Differential Privacy: Does it actually work?
18:10 – Myth #3: The Problem with AI Risk Taxonomies
21:30 – Building an Interdisciplinary Risk Radar
24:00 – Myth #4: "We Red-Teamed Once, We're Fine"
27:50 – Iterative Security: Threat Modeling with STRIDE & PLOT4AI
31:15 – Myth #5: The Next Model Version Will Fix This
34:20 – The Case for Local LLMs & Diversifying Providers
38:05 – Q&A: Training Your Own Security Routers
🔗 Transcript & slides available on InfoQ: bit.ly/4sFRhUf
#AIprivacy #AIsecurity #LLMSecurity #GenerativeAI #CyberSecurity #EngineeringLeadership #InfoQ
📅 Subscribe for weekly talks from senior engineers at companies like Netflix, Spotify, and IBM.
As AI moves from augmentation to full automation, the stakes for privacy and security have never been higher. In this InfoQ video, privacy expert Katharine Jarmul breaks down why current AI "guardrails" are easier to bypass than you think and why senior engineers cannot rely on model providers to solve privacy for them.
Learn the architectural realities of overparameterization, why models "memorize" sensitive data, and how to build a culture of psychological safety to catch incidents before they scale.
⏱️ Video Timestamps (For Navigation)
0:00 – The Shift: From Augmentation to Automation
2:15 – The Problem with Privacy "Fearmongering"
4:30 – Myth #1: Why Guardrails (Software & Algorithmic) Fail
7:45 – Bypassing Filters: Variable Renaming & ArtPrompt Attacks
10:12 – Myth #2: Overparameterization & The Death of Overfitting
13:20 – Can LLMs Learn Without Memorizing? (The Data Leak Risk)
15:45 – Differential Privacy: Does it actually work?
18:10 – Myth #3: The Problem with AI Risk Taxonomies
21:30 – Building an Interdisciplinary Risk Radar
24:00 – Myth #4: "We Red-Teamed Once, We're Fine"
27:50 – Iterative Security: Threat Modeling with STRIDE & PLOT4AI
31:15 – Myth #5: The Next Model Version Will Fix This
34:20 – The Case for Local LLMs & Diversifying Providers
38:05 – Q&A: Training Your Own Security Routers
🔗 Transcript & slides available on InfoQ: bit.ly/4sFRhUf
#AIprivacy #AIsecurity #LLMSecurity #GenerativeAI #CyberSecurity #EngineeringLeadership #InfoQ
📅 Subscribe for weekly talks from senior engineers at companies like Netflix, Spotify, and IBM.
