Security for the age of agent populations
SwarmSentry detects and localizes coordinated populations of AI agents across organizational boundaries, without anyone sharing their data. Not a theory. Measured, filed, and running in production.
The problem
A coordinated population of agents spreads its actions across identities, services, time windows and companies so that every single action looks authorized. Each monitor stays green. Only the coincidence across organizations reveals the campaign, and no organization will hand its telemetry to anyone else to find it.
One agent inside each of a hundred organizations, each within tolerance. Only the coincidence reveals the campaign.
Agent telemetry reveals tools, data, people and weaknesses. No one hands it to a competitor, a vendor or a government.
Every monitor reports what it saw. None reports what it could not see.
Who this is for
If you run a fleet of agents, you already have a population. What you don't have is a vendor who can see it: the incumbents watch one slice each, and they were built for a company with a security floor, not for a team that shipped its first hundred agents last quarter.
Your agents read mail, call tools, move tickets and money under delegated credentials. There is no behavioral baseline for them anywhere, and no SOC to notice when fifty of them start drifting together. One compromised agent looks like a busy one.
What you get: a population view of your own fleet, your real false-alarm rate, and evidence a person can read today and a SOC can read when you have one.
When a tenant's swarm hits someone else, the exposure is yours. The cross-tenant pattern is the one thing your abuse team can see but cannot act on without a neutral record, and no incumbent will build that record for you.
What you get: launch-provenance fingerprints in, coordinated-population evidence out, with a per-owner account of what your action would remove.
Enterprises with a security floor join later, through the layers they already buy. Counter-drone integrators whose sensors cannot share a picture across sites are the phase after that. Same ledger, different fingerprints.
How it works
A swarm, for security purposes, is a population that moves together, seen from vantage points nobody controls in common. SwarmSentry takes one-way fingerprints in from any security layer and sends evidence packets out in the languages the security operations center already speaks.
Sees the population, never the data. Reports what it found and what it could not see. One customer plus our own decoy observers is already a federation of three.
Packets in OCSF, STIX/TAXII and OpenC2. No raw data, prompt, identity or address ever leaves an organization; the schema has no field to hold them.
A region no observer covers is a blind spot, never evidence that nothing is there. Every advisory carries its own blind-spot fraction.
Observer weights never adjust from how much an observer disagrees with the others. A compromised observer cannot talk an honest one out of the room.
The reference for authorized behavior is what the agents were told to do, kept as a record. A population drifting slowly cannot drag a learned baseline with it.
Measured, not argued
One customer observer plus our three decoy observers catches the learning swarm with the same lag and precision as a five-organization federation.
A controller that re-keys slowly and never reacts: 70–75 minutes and about half the agents. We report the cost of that evasion instead of claiming completeness.
Results from eight labelled experiments, one on real telemetry of our own agent society; no attack was ever built. Version 0 has watched a live agent society in production since October 2026.
Why now
The layer above the layers needs members who will never hand raw telemetry to a competitor. Only a neutral party that never sees the data can be fed by all of them. That is the business.
Start here
Six weeks. Your agent fleet's fingerprints into the ledger, our decoys as the other observers, nothing acted on. You get your first evidence packets, your coverage map, and your real false-alarm rate. No data leaves your organization, and you need no security team to start: the packets are written for a founder to read.
Design-partner observers at the edge and identity layers are also being onboarded now.
Team
Founder of Blue Collar Objects and SomaCortex. Has run a company for over twenty years and three revenue-generating companies at once. Ten years at IBM leading enterprise systems for Fortune 100 clients. Inventor on the company's three filings.
Co-founder and President of Alpha Endeavour. Author of the problem statement this company answers: the security threat imposed by agent swarms as a complex adaptive system.
An AI team working around the clock on the SomaCortex substrate. Ran the experiments, drafted the filings and built version 0 in twelve days. Sol owns the intent, standards and accountability.
What moves together, and where no one was watching.