I Abuse Your LLM Live with 10 Security Vulnerabilities
summary
A live demo shows ten concrete failure modes of large language models, including prompt injection that reveals hidden system prompts, agent tool calls that execute SQL and leak user data, and LLM outputs that can trigger XSS. It highlights supply-chain risks from compromised models and datasets (e.g., “King Face” scans) with expired-domain link rates of 0.3–6.5% and data poisoning around 0.1%, and presents mitigations such as lightweight LMs as input/output filters, function-call logging, sandboxing with human approval, anonymized indexing, whitelists, quotas, and strict output filtering.
AI-written summary · about
chapters (13)
- 0:00LLM security intro
- 2:56Live vulnerable LLM demo
- 7:01Prompt injection explained
- 10:43Function call risks
- 16:13Sensitive data leaks
- 19:23Excessive agent permissions
- 21:15Supply chain vulnerabilities
- 25:51Model data poisoning
- 30:12Misinformation and hallucinations
- 33:40Embedding manipulation risks
- 36:10XSS via LLM output
- 39:53Resource exhaustion
- 42:26Best practices
related talks
50mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges
20mSecurity Threats in Modern LLM Applications
53mBreaching LLM-Powered Applications: Overcoming Security and Privacy Challenges
23mSecurity and auditing tools in Large Language Models (LLM)
49mBreaching LLM-Powered Applications: Overcoming Security and Privacy Challenges
28mSecuring LLMs: Insights into OWASP top 10
36mCompromising LLMs: The Advent of AI Malware
31mHacking with Words...