Doors of (AI)pportunity: The Front and Backdoors of LLMs
summary
Claim: local LLMs and ML models are exploitable quickly on a CPU — a jailbreak or permanent backdoor in minutes. Demonstrations include a Bobby Tables-style KROP injection that produces DROP TABLE, a multipart image-to-code exploit that yields a reverse shell, repository sideloading that plants a computational graph backdoor to flip numpy.loads to allow pickle, and an automated gadget hunter called autocrop plus a universal policy puppetry prompt that overrides refusals by ablating a single vector.
AI-written summary · about
chapters (10)
- 0:00Introduction to AI doors
- 1:30Understanding prompt injection
- 2:52Knowledge return oriented prompting
- 6:27Automating prompt attacks
- 11:30Policy puppetry and alignment
- 17:27Indirect prompt injections
- 21:52Responsible local AI testing
- 26:59Computational graph backdoors
- 35:19Graph-based trigger injection
- 43:58Advanced evasion and conclusion
related talks
53mDoors of (AI)pportunity: The Front and Backdoors of LLMs
36mCompromising LLMs: The Advent of AI Malware
32mFuture Open Source LLM Kill Chains
09mLLM hacking is underrated
39mLLMs everywhere: The future is now and the past keeps repeating itself
53mBreaching LLM-Powered Applications: Overcoming Security and Privacy Challenges
50mSecuring LLM-Powered Applications: Overcoming Security and Privacy Challenges
44mI Abuse Your LLM Live with 10 Security Vulnerabilities