Compromising LLMs: The Advent of AI Malware
summary
Claim: prompt injections enable a new class of malware that runs entirely inside LLMs and can compromise many instances at once. Demonstrations include a website that exfiltrates user names via Bing and use of markdown image fetches to leak data; plugins and browsing lead to account data leaks; code-interpreter VMs are jailbreakable; Copilot autocomplete is poisonable; prototypes show indirect prompt injection enabling remote-control C2 and persistent memory reinfection. Paper released Feb/May, disclosures filed with OpenAI, Microsoft and the German BSI; 100 million+ user integrations create scale risk.
AI-written summary · about
chapters (11)
- 0:00AI malware threats
- 1:30LLM security threats evolution
- 3:40Prompt injection attacks
- 8:13Indirect prompt injection
- 11:49Bing Chat exploits
- 17:30ChatGPT plugin risks
- 20:37Code interpreter risks
- 23:44Future attack vectors
- 26:59Military AI risks
- 30:47Why mitigations fail
- 35:23Key takeaways
related talks
39mLLMs everywhere: The future is now and the past keeps repeating itself
44mI Abuse Your LLM Live with 10 Security Vulnerabilities
26mWhen AI Gets Manipulated: Security Risks in LLM Applications
27mSecuring The Gateway And Mitigating Risks In LLM API Integration
50mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges
01h 01mIntroduction to AI Security
49mDoors of (AI)pportunity: The Front and Backdoors of LLMs
20mUnderstanding and Mitigating Risks in LLM-Powered Web Apps