Preventing an OWASP Top 10 in the world of AI
summary
Claims a working exploit of prompt injection where a chatbot learns a provided filename like malicious.exe and later serves it back as an action item, and documents a separate vulnerability where API response types are mutated allowing GraphQL style manipulation (AI Dungeon example). Finds near zero logging or monitoring, broken authentication, and recommends human-in-the-loop workflows plus input and file standardization and logging. A live replica demo fails due to maintenance but mitigations are implemented.
AI-written summary · about
chapters (6)
- 0:00Start
- 0:20Owasp Chair for Columbus Ohio
- 10:35Prompt Injection
- 17:39Integration Strategies
- 18:03Human in the Loop
- 22:21Key Takeaways
related talks
54mA Practical Guide to AI Threats & Mitigation Strategies
01h 14mBreaking the Bot: GenAI Web App Attack Surface & Exploitation
18mDefending the unknown using AI and cybersecurity
28mSecuring LLMs: Insights into OWASP top 10
36mAgentic AI Under Attack: Live Demos of Exploits Through Autonomy and Trust
57mAI Applications Security Puzzle
24mSecuring Language Models Against Prompt Injection
51mPrompt Injection: When Hackers Befriend Your AI