Ethically Hacking LLMs, AI agents - what threat actors can do to your training data that powers your agents to execute code remotely.

My talk will attempt to spell out what Machine Learning LLMs are, AI agents - what lies at the nucleus of these first. I will then show how they work, what threat actors can do to exploit them and cause data integrity issues, even compromise them away from what they were meant to achieve.

I will demonstrate through code, using Python - its library Pickle and a select method, how threat actors can abuse serialization and deserialization to do harm. How it is possible to conduct SQLi within the RAG eco-system by malicious actors.

If time permit, a brief look at Model Context Protocol and possible abuse of this.

A few pointers on mitigation will also be shared with the audience.