Hidden prompt injection
Prompt-injection text is hidden inside the logs to try to trick the assistant.
I build a test cloud setup with Terraform, run safe, simulated attacks against it, and send all the security logs into a data pipeline I built. Detection rules written in SQL flag the attacks, and an AI assistant writes a short summary of each alert. Then I attack the AI assistant too, and add defenses so it can't be tricked.
Section 01: Section 01
A test cloud setup is built with Terraform.
Safe, simulated attacks are run against it.
All the security logs flow into a data pipeline I built.
Detection rules written in SQL flag the attacks.
An AI assistant writes a short summary of each alert.
Each AI alert summary covers three things
Triage output
Section 02: Section 02
The assistant reads the logs, so the logs become a way to attack it.
Prompt-injection text is hidden inside the logs to try to trick the assistant.
Defenses are added so the assistant is harder to trick.
Section 03: Section 03
Problem one
Security teams drown in cloud logs.
Problem two
AI tools that help them can be turned against them.
This project works on both problems.
Section 04: Section 04
The project repository is on GitHub. The write-up is coming soon.