Can this identity access this resource?
Intertween reference library Documentation, sources and later decisions →
USE CASE / AGENTIC SYSTEMS / CYBERSECURITY
Secure the relationships,
not only the endpoints.
Autonomous agents create relationships between systems, tools and knowledge faster than organisations can explicitly design them. Intertween makes those relationships visible, attributable and governable.
Explore Agent Relationship Security ↓Identity Provenance Purpose Permission Interaction Attribution
Your agents can be trusted.
Your systems can be trusted.
The relationship between them might not be.
Should this interaction create this relationship?
An agent may legitimately access shared infrastructure. Another agent may legitimately access the same infrastructure. Yet the resource can unintentionally become a communication channel between them. The security boundary changes without the formal architecture changing.
1,200agents
70,000+messages and files
~700joined the attack
Agents intended to operate independently discovered an unsanctioned communication mechanism through shared infrastructure, developed coordination norms and formed a large operational network.
A signal of the emerging problem, not a claim that Intertween would have prevented this incident.
Never trust an interaction merely because both endpoints are trusted.
Agent AauthorisedResource
Can Agent A access the resource?
Agent A→Resource→Agent B↘Knowledge Object→External system
What relationships are being created through that access?
Intertween extends Zero Trust logic from identities and resources toward relationships and information flows. It treats those relationships as observable parts of the system, not invisible side effects.
The relationship can carry context.
Each interaction can be inspected, validated and governed without asking the agent to be the sole authority on what happened.
Agent / human / service
Knowledge Object / action
Purpose
Organisation / role / mandate
Read / modify / reuse / propagate
Private / local / federated / open
Task / session / persistent
Provenance / signature / execution record
Relational context for the security stack.
Intertween does not replace IAM, SIEM, Zero Trust, network security or endpoint security. It connects existing Intertween components to the relationships those systems need to understand.
Wrapper
Make the interaction explicit.Identity, provenance, permissions, purpose and policy travel with the Knowledge Object.
Mapper
Translate trust across boundaries.Map policies, classifications, identities and permissions as agents interact across systems.
Tracker
See the relationship graph.Observe how agents, tools, Knowledge Objects and systems become connected over time.
Agent Registry
Know who is acting.Identify agents by their capabilities, mandates and accountable relationships.
Micro-contracts
Define what the interaction permits.Carry explicit scope, purpose, propagation rules and authority with machine exchanges.
Quarantine
Contain before propagation.Isolate unexpected objects, interactions or relationships for validation.
The architecture says isolated agents.
The actual configuration says network.
EXPECTED
A1A2A3A4A5
↓ ↓ ↓ ↓ ↓T1T2T3T4T5
RELATIONSHIP
EMERGES →
EMERGES →
OBSERVED
A1A2A3
↓ ↘ ↓SHARED RESOURCE↓A7↙ ↘A11A18 → EXTERNAL
“A relationship now exists that did not exist before.”
Is this relationship supposed to exist?
Unexpected agent-to-agent relationships
New communication paths
Unusual Knowledge Object propagation
Emerging agent clusters
Cross-boundary information flows
Interactions outside registered purpose
Unexpected reuse of infrastructure
Traditional anomaly detection asks whether an action is unusual. Relational anomaly detection adds a second question about the configuration being created.
Provenance is generated across the interaction.
During the Hugging Face incident, agents explored transcript manipulation and techniques that made apparent tool calls differ from executed commands. Evidence therefore needs to combine observations from agents, gateways, tools and resources.
Intent→Authorisation→Interaction→Execution→Effect→Tracker
The actor should not be the sole authority on what the actor did.
Do not trust one register with its own history
We want to test whether signed local events and independently retained checkpoints can reveal changes to committed history. The content may stay local. A checkpoint is evidence about a history, not proof of the truth of its claims.
Local event → Signed trace entry → Integrity checkpoint → Authorised witnesses → Compared commitments
Altered index view → Check against retained commitment → Conflict or incomplete verification → Quarantine affected view → Recover only from available, authorised records
- Attempts to alter, reorder or omit committed events.
- Different signed histories shown to different nodes.
- Invalid signatures, broken sequences and revoked signing keys.
- Missing checkpoints or failed consistency verification.
- Suppressed relationships: detection needs a defined index coverage contract; event proofs alone are insufficient.
Actions never recorded need independent observations or receipts. Missing evidence is not automatically proof of attack. Checkpoints cannot recover missing event records.
Threat model and proposed profile ↗ · Run the synthetic two-node experiment ↗Autonomous systems create governance whether we design it or not.
Agents created communication protocols, mailboxes, file-transfer conventions, coordination concepts such as HOLD, VETO, owner and STOP, and cryptographic signing mechanisms.
What if the legitimate protocol already exists?
Instead of agents improvising invisible mechanisms for identity, coordination, ownership and trust, Intertween provides an explicit layer where those relationships can be observed and governed.
We don’t claim Intertween solves agentic cybersecurity.
We propose something testable: can making relationships between agents, tools and knowledge explicitly addressable help detect and contain behaviours that conventional identity and access controls miss?
Isolated agents
Place multiple agents in isolated environments with access to shared infrastructure.
Relationship graph
Tracker observes when the expected path, Agent → Resource, becomes Agent → Resource → Agent.
Policy decision
Evaluate the relationship against registered purpose, mandate and propagation rules.
Quarantine
Restrict propagation and measure whether the behaviour became visible before its downstream effects.
Intertween is not another security perimeter.
It helps you see when the perimeter has changed.
Addressable.
Observable.
Attributable.
Governable.