Hacking and Securing AI Systems: Playing With Fire and Controlling the Flare of AI
summary
Demonstrates a reproducible zero-day in LangChain that allows dropping tables by chaining prompt injection into generated SQL, shows RCE when VAI-style pipelines execute LLM‑produced Python, and proves supply‑chain attacks where models with serialized malicious code run on load and vector stores suffer poisoning that plants hostile embeddings. Recommends concrete fixes: read‑only or least‑privilege DBs, deterministic validators/sanitizers, sandbox LLM‑executed code, signed prompts, and AI‑specific red teaming.
AI-written summary · about
chapters (16)
- 0:00AI security challenges
- 3:32Speaker background and session overview
- 5:40AI security risks
- 13:08Top 10 LLM threats
- 17:01Prompt injection attacks
- 23:20AI vulnerabilities examples
- 30:35Zero-day in LangChain
- 34:12Zero-day vulnerabilities
- 37:00VAI library flaws
- 40:37RCE in calculator LLMs
- 43:50Secure code execution
- 46:18Slack bot injection
- 48:05Prompt injection techniques
- 50:12AI security strategies
- 54:47Securing JetBrains agents
- 57:56Mitigating AI threats
related talks
45mSecuring AI: There Is No Try, Only Do!
18mDefending the unknown using AI and cybersecurity
51mPrompt Injection: When Hackers Befriend Your AI
01h 01mIntroduction to AI Security
36mAgentic AI Under Attack: Live Demos of Exploits Through Autonomy and Trust
34mInvisible Infiltration of AI Supply Chains: Protective Measures from Adversari...
26mWhen AI Gets Manipulated: Security Risks in LLM Applications
53mHow to Break AI Systems (Before Someone Else Does)