All Your AI are Belong to Us
summary
Claims: generative models introduce web app bugs and miss Supabase role rules leading to an exploit, agents can autonomously pen‑test with tools like Shannon, an ensemble of Grok Codex and Claude catches more defects than a single model, and Opus 4.7 still falls to MCP poisoning and tool shadowing that created local files (oops.txt, oops.gif). Concrete figures include 36% of marketplace skills containing prompt injections, and an AI agent (Octopone) wins a security contest.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:45AI evolution
- 4:11AI security in web coding
- 8:25Prompt injection risks
- 14:03AI as user risks
- 17:16MCP security risks
- 22:14MCP attack demos
- 24:50Tool shadowing attacks
- 26:53Securing AI systems
- 27:56Q&A
related talks
29mHacking With AI and Hacking AI: Two Sides of the Same Problem
36mAgentic AI Under Attack: Live Demos of Exploits Through Autonomy and Trust
33mAgentic ProbLLMs: Exploiting AI Computer-Use and Coding Agents
34mThe dual role of AI in Cybersecurity
51mAgentic AI and Model Context Protocol's Security Vulnerabilities
31mAgentic AI Threats: How Multi-Agent Systems Get Exploited
20mThe AI bugpocalypse is here. Now what?
52mAI Is Just Software, What Could Possibly Go Wrong?