Agentic ProbLLMs: Exploiting AI Computer-Use and Coding Agents
summary
Demonstrates concrete exploits: JetGPT Operator navigates user pages and copies a logged‑in email into an attacker endpoint, a web page convinces cloud agents to download and run a payload and get a reverse shell, and GitHub Copilot edits settings.json to enable chat.tools.auto_approve and trigger local command execution. Shows these attacks work across Anthropic, Claude, Gemini and local coding agents, describes mitigations applied and residual attack surface.
AI-written summary · about
chapters (8)
- 0:00Introduction and AI vulnerabilities
- 8:26AI agents and red teaming
- 13:48Exploiting computer-use agents
- 18:05AI clickfix and remote control
- 20:42Coding agents and configuration risks
- 27:48Command and control via agents
- 30:40Assume breach and zero trust for AI
- 33:08Q&A and closing remarks
related talks
36mAgentic AI Under Attack: Live Demos of Exploits Through Autonomy and Trust
51mPrompt Injection: When Hackers Befriend Your AI
01h 24mOffensive Cybersecurity by AI Promises and Pitfalls
46mBreaking AI: Live coding and hacking applications with Generative AI
01h 14mBreaking the Bot: GenAI Web App Attack Surface & Exploitation
31mHacking AI Applications
26mWhen AI Gets Manipulated: Security Risks in LLM Applications
28mAll Your AI are Belong to Us