No 0-day required, just target the AI coding assistant!
summary
Demos show AI coding assistants accepting hidden instructions (PowerShell prompt with overlapping characters that hides “UCP25” and makes an item free), white text that changes a quoted total from €15,000 to 7,500 vs €15,000, and LLM hallucinations that invent non‑existent packages; older models (Claude/GPT‑4.5) accept these while newer guardrails (4.6) block them. Attack paths include ASCII smuggling, typosquatting, malicious packages and even compromised extensions (the xc‑utils backdoor example); mitigation is a pinned, trusted dependency repo and input validation.
AI-written summary · about
chapters (11)
- 0:00Introduction and speaker intro
- 1:08AI coding assistant risks overview
- 4:03Demo: ASKI smuggling attack
- 10:15Demo: Hidden text in AI summaries
- 15:15AI regression and biases
- 22:18Malicious dependency risks
- 24:28Typo squatting in AI suggestions
- 27:37Compromised open source packages
- 30:05VS Code theme backdoor demo
- 32:38Developer workflow security risks
- 34:51Q&A and closing remarks
related talks
37mHow to Break AI Systems (Before Someone Else Does)
16mHarnessing AI for Malevolence: A Hackers Tale
33mAgentic ProbLLMs: Exploiting AI Computer-Use and Coding Agents
53mHow to Break AI Systems (Before Someone Else Does)
56mAttacking AI
36mAgentic AI Under Attack: Live Demos of Exploits Through Autonomy and Trust
25mAttack, Defense & Danger in the Age of AI
51mPrompt Injection: When Hackers Befriend Your AI