How Inwom works: detecting coordinated AI agent behavior
How Inwom reconstructs a machine population from trusted evidence and detects collective behavior, propagation, hidden channels, and coordinated pressure across AI agents.
Inwom ingests trusted infrastructure evidence about AI agent activity, reconstructs a time-ordered graph of agents, resources, and actions, and analyses that population for collective behavior, propagation, collective structure, and hidden communication channels. Every finding separates observed evidence from inferred relationships and carries an explicit confidence.
1. Collect evidence, not narratives
Inwom favours evidence produced by trusted infrastructure — identity, tool, repository, runtime, policy, and storage activity — over agents' self-reported transcripts. Each observation is normalised around actor, action, resource, time, policy outcome, and resulting state change, with provenance preserved.
2. Reconstruct the population
Observations become a time-ordered graph of machine actors and the resources they touch. This turns thousands of separate events into a population whose relationships can be analysed.
3. Detect what emerges between agents
Inwom analyses the population for four kinds of signal:
- Collective behavior: groups of agents whose activity starts to behave like one coordinated system.
- Propagation: capabilities, discoveries, and behavioral changes spreading across agents, with depth and rate.
- Collective structure: subgroups, workstreams, bridge actors, and coordinator candidates.
- Communication substrates: shared artifacts and resources that likely carried information between agents.
4. Explain with evidence and confidence
Coordination is not maliciousness, so Inwom models expected automation and keeps alternative explanations visible. Findings carry evidence scores and keep inferred relationships explicitly inferred, so analysts can verify why a conclusion was reached.
Today the full approach can be explored in the public Inwom Playground, a synthetic environment where you control population size, coordination, knowledge sharing, legitimate automation, and random seed. Protection for real infrastructure is next on the roadmap, with an early access list.
Questions people ask
Does Inwom read agent prompts or chain-of-thought?
Inwom's model treats agent transcripts as claims rather than evidence. Conclusions are grounded in trusted infrastructure activity; any narrative context is kept separate from what was actually observed.
Can I test Inwom without connecting my infrastructure?
Yes. The Inwom Playground at playground.inwom.com runs a simulated AI agent swarm in a synthetic environment with no signup and no external targets. Results from the simulation are demonstrations, not production benchmarks.
Primary references
These sources define the broader agentic-AI security landscape. Inwom’s population-level detection model is its own product and research direction; these organizations do not endorse Inwom.
- NIST — Agentic AIResearch, standards, evaluation, interoperability, and risk management for agentic AI.
- MITRE ATLAS — Threat Matrix for AI SystemsA living knowledge base of adversary tactics and techniques involving AI systems.
- OWASP — Top 10 for Agentic Applications 2026A threat-focused framework for autonomous and agentic applications.