Research
Defending at machine speed.
Security teams already handle automation. What changes with AI agents is scale: agents multiply the number of actions, delegations, and tool boundaries a defender has to evaluate — often faster than a person can review them. Threeum Defense is our work on that problem.
The thesis
When many agents act across many identities and systems, individually harmless actions can add up to one coordinated effort. Catching that means reasoning about the behavior, relationships, and likely next actions of automated systems — not just individual files, signatures, or events.
Our approach pairs fast, deterministic checks with adaptive models, and keeps authority deterministic: models raise or lower suspicion, but your policy decides what happens.
Scope of results
This is early work. There are no published benchmarks yet, and the evaluation harness and frozen held-out datasets are still being built. Simulated event streams are labeled as simulation and are not presented as real-world evidence.
When results are published, they will state the dataset, its provenance and license, the evaluation split, and the exact scope of each claim.
Open directions
Where the hard problems are.
Swarm correlation
Linking distributed, low-signal actions into one campaign with an explainable confidence, using temporal and graph methods.
Agent behavior
Recognizing agents by behavior and detecting instruction drift and likely prompt injection — without over-blocking legitimate automation.
Edge efficiency
How small a specialist model can be while remaining useful for continuous, low-overhead detection.
Gated learning
Turning confirmed incidents into reusable signatures under regression gates and human review — never self-promotion.
Explainability
Every detection should answer: what happened, why it's suspicious, which signals contributed, and what to do.
Privacy-preserving sharing
Sharing threat intelligence across nodes without sharing sensitive underlying data.
How we evaluate
Reproducible seeds; chronological or grouped splits; validation-selected thresholds; and frozen, held-out public data for leak-free evaluation. We measure precision, recall, F1, false-positive and false-negative rates, detection and inference latency, and resource use — and we report what we could not measure.
Research notes live in the repository under research/, with per-family model cards under ml/ as artifacts are produced.
Primary-source references
Third-party standards and taxonomies we study. Links are external; inclusion is not endorsement by those parties.