Who to blame? The AI, The Programmer, or The Prompt?
summary
Five live case studies show concrete bugs and fixes: a Copilot image proxy leak that exfiltrates secrets via camo URLs and leads to CVE-2025-53773, an unauthenticated Olama Docker API that enables manifest poisoning and local-to-remote RCE, AWS MCP string concatenation that becomes shell injection and long-lived reverse shells, a prompt‑injection path that exposes private repos, and a Microsoft zero‑click exfiltration chain CVE-2025-32711. Each case gives the exploit vector, the patch applied, and who changed behavior.
AI-written summary · about
chapters (12)
- 0:00AI blaming trial introduction
- 3:48Camo attack on GitHub Copilot
- 8:54Defending GitHub Copilot
- 11:25Olama’s local flaw
- 14:19Defending Olama
- 16:51AWS MCP shell injection
- 19:50Defending AWS and Amazon
- 21:34Invisible collaborator threat
- 25:09AI security mental model gap
- 30:09Responsibility for AI vulnerabilities
- 32:35Closing remarks
- 33:00Q&A
related talks
51mPrompt Injection: When Hackers Befriend Your AI
56mAttacking AI
43mThe AI Assistant’s Betrayal: One-Click for AI to Turn into the Perfect Insider
36mAgentic AI Under Attack: Live Demos of Exploits Through Autonomy and Trust
45mPrompt-Jacking: The Rise of a New Supply Chain Risk
51mFrom Text to Flaws: vulnerabilities in applications with Generative AI and LLMs
47mToolShell, Patch Bypass, and the AI That Might Have Seen It Coming
46mNot My Vibe: When AI Coding Agents Go Off the Rails