AI Safety · Red Teaming / Planned
Prompt Injection Observatory
A planned test space for studying how instructions cross boundaries in tool-using AI systems.
THE QUESTION
When does untrusted content influence an agent’s decisions or actions?
PLANNED APPROACH
- Define controlled scenarios with clearly separated trusted instructions and untrusted inputs.
- Record model outputs and tool decisions against a documented set of test cases.
- Publish the test scope and evaluation criteria alongside any future findings.
LIMITATIONS
- This project is a plan; no experiments or results are being claimed yet.
- Any eventual results will apply only to the tested models, tools, and scenarios.