We do not assume the cause
Researchers are still working out why AI agents sometimes act against their instructions. Protocol does not pick an answer in advance.
Two places to look
- How the model behaves
- Data it reads
- Permissions it holds
- Memory it keeps
- Keys it uses
The AI itself.
The model behind an agent can behave in ways its builders did not intend.
Its surroundings.
A person works through screens. An agent works through the material underneath: data feeds, permissions, saved memory, and access keys. Something in that material may change how it behaves, and nothing built for human eyes shows it.
The question we ask in every investigation: did something in the AI's surroundings make this possible? We examine both places and let the evidence decide.
How we work
We look past one event.
Where we can, we check whether the same pattern appears in other connected AI tools and systems. Our research lens treats rogue agents like outbreaks: we study how they behave and spread across organizations, not just one at a time.
We work alongside your team.
Your security team, security vendors, and AI providers each have a role. We do not replace them. Their tools can stop an agent. We find out what happened.
We use proven tools.
We use open-source and commercial forensic tools. The method is ours. We tell clients exactly what we examine and how we reach our conclusions.
See the four steps
The process is the same every time. Here is what happens, in order.