Securing AI: There Is No Try, Only Do!
summary
LLMs suffer concrete failures: prompt‑injection and optimization attacks let adversaries override system instructions; pretrained data from Common Crawl and small poisons create backdoors (example magic trigger Reclude Obscurum that disables guardrails); fine‑tuning regresses safety (a Mistral‑7B fine‑tune turns helpfulness into actionable wrongdoing); RAG and vector DBs leak or reverse embeddings and are trivially poisoned; an AI firewall test shows ChatGPT decodes obfuscated SSNs, and tool‑calling models succeed only about 78% of the time.
AI-written summary · about
chapters (15)
- 0:00AI security concerns
- 1:09What is AI and security
- 4:45Attacks on ML systems
- 6:54Why LLMs differ
- 10:27LLMs as parrots
- 12:14Prompt injections
- 16:07De-alignment risks
- 17:51Hallucinations and leaks
- 19:56Data poisoning
- 23:52AppSec for AI systems
- 28:05Retrieval augmented generation security
- 32:41AI agent security
- 35:55Limitations of traditional tools
- 38:07AI security importance
- 40:14Q&A
related talks
35mHacking AI Systems: How to (Still) Trick Artificial Intelligence
01h 01mIntroduction to AI Security
18mDefending the unknown using AI and cybersecurity
57mAI Applications Security Puzzle
01h 24mOffensive Cybersecurity by AI Promises and Pitfalls
53mHow to Break AI Systems (Before Someone Else Does)
01h 05mNavigating the Security and Privacy Landscape of Modern AI
53mHow AI is impacting cybersecurity. Does it?