NeuronX Fuse
ActiveAgentic security research. AI agents read email, call tools, and touch company data. Fuse studies how they get attacked, and how they can defend themselves.
The questions
What we're trying to answer.
- 01 How do agents resist prompt injection?
- 02 How do we stop one agent from leaking another company's data?
- 03 Can defenses act as fast as the attacks?
- 04 How should an agent ask a person before it acts?
Our approach
How we work on it.
Red-team our own agents.
We attack the agents we build before anyone else does, and keep the cases that break them.
Gates before actions.
Every message passes checks for spam and injected instructions before an agent acts on it.
Least access by default.
An agent sees only the data its company and its role allow. Nothing more.
People stay in the loop.
When an agent isn't sure, it stops and asks a person.
Where it’s going
Sovereign and high-stakes use. We want Fuse to protect organizations that can’t send data outside their walls. That’s a goal, not a product we sell today.
Questions
Straight answers.
- Is Fuse a product?
- No. It's a research program. Its results ship inside AutoScale.
- Does NeuronX hold security certifications?
- No. We'll say so plainly until that changes.
Collaborate
Research agent security with us.
Professors, labs, and companies can propose a joint project. Tell us what you want to study together.
Choose Fuse as the program.