AI Security Crisis: Jailbreaks, Prompt Injection & How to Protect Your Agents

Agentic AI Institute · Intermediate ·📰 AI News & Updates ·4w ago
Sign up to get my learning resources: https://forms.gle/sRNjXnsurNxNAUQW7 2026 was predicted to be the year of agentic AI moving into enterprise production. But there’s one problem: - AI agents are failing publicly. - Jailbreaks are succeeding. - Prompt injection is real. - Trust is eroding. In this session, we break down the real AI security crisis and what product managers, founders, and builders must do before shipping agents. You’ll learn: • Why jailbreaking is an arms race • What prompt injection really is (and why it’s dangerous) • DeepSeek’s 100% jailbreak success case • Devon AI se…
Watch on YouTube ↗ (saves to browser)

Chapters (16)

2026: The year of agentic AI… but trust is breaking
3:20 Why AI security failures are costing real money
6:10 Jailbreaking explained (DAN attack & DeepSeek case)
11:30 Why performance ≠ security
14:00 Prompt injection explained (and why it’s worse)
18:30 Devon AI security failure case study
23:40 OpenClaw risks and real exploit paths
28:20 Why AI security is structurally hard
33:00 Why probabilistic guardrails fail
37:10 The 3 remedies: Architecture, Red Teaming, AI SecOps
40:00 KEL architecture (Dual LLM separation model)
46:30 Red teaming tools (Microsoft, Nvidia, DeepTeam)
49:30 AI SecOps: Monitoring, lifecycle, governance
54:00 Live demo: Attacking an AI agent using Azure
59:30 How jailbreak prompts bypass guardrails
1:04:00 Reviewing atta
What happens if you eat a silica gel packet? - Vivian Jiang
Next Up
What happens if you eat a silica gel packet? - Vivian Jiang
TED-Ed