Using GenAI on and inside your code, what could possibly go wrong?
summary
Academic benchmarks find roughly 40% of generated apps contain vulnerabilities; a Veracode-style dataset shows ~71% of orgs carry year‑old security debt and 45% have critical issues. Tests across models (GPT‑4, StarCoder, Claude) and tools (Copilot, Cloud Code, Git MCP, BoxBench) show security improves with stricter prompts. A live Semantic Kernel demo reading /etc/hosts exposes plugin exfiltration risks. Platform risks include jailbreaks, supply‑chain poisoning where 1% training‑data control influences outputs, and huge compute footprints (estimated $40M to train GPT‑4).
AI-written summary · about
chapters (9)
- 0:00GenAI and conference overview
- 1:53How LLMs work and risks
- 9:32AI tools in development
- 17:35Security risks of AI code
- 29:34Secure LLM integration
- 36:43Prompt injection vulnerabilities
- 43:26LLM platform risks
- 50:06Practical AI security
- 52:55LLMs in security and review
related talks
52mUsing GenAI on and inside your code, what could possibly go wrong?
12mTrusting AI with Code: How Secure is AI-written code?
31mHacking AI Applications
30mWhat Are the Security Risks Associated with AI-Generated Code? 🤔
41mDon’t Make This Mistake: Painful Learnings of Applying AI in Security
27mSafeguarding Generative AI Applications: Navigating Security Challenges and Best Pract...
53mHow AI is impacting cybersecurity. Does it?
46mBreaking AI: Live coding and hacking applications with Generative AI