Easycoreby Easylab AI
Request a demo

Security is enforced at runtime.

Most AI security policies are documents. Ours run in production: every sensitive action is arbitrated the moment it goes out, by a defence in depth made of several independent layers.

One agent, one machine.

Each agent runs on its own dedicated machine. Isolation is physical before it is software: even if compromised, an agent has no path to another client's data.

InternetExposed bastionthe only exposed zoneWeb cockpitPublic APIAuthenticationMetadata cataloguePrivate hubprivate network onlyOrchestratorSealed audit chainTrust scoreIsolated agentsone agent, one machineAgent AAgent BAgent Cmutual authenticationInternetExposed bastionthe only exposed zoneWeb cockpitPublic APIAuthenticationMetadata cataloguePrivate hubprivate network onlyOrchestratorSealed audit chainTrust scoreIsolated agentsone agent, one machineAgent AAgent BAgent Cmutual authentication
Principle diagram. The detailed plan is in the file handed over in due diligence.
Machine
A dedicated machine per agent.
Resources
Each agent's own volumes, networks and collections.
Scope
Data, memory and runtime isolated at infrastructure level.
External content
Emails, web pages, third-party files: external content is framed as data, never as instruction. Mail is handled by a worker with restricted tools.

Nothing travels in clear.

Conversations kept by the platform are encrypted on disk, in authenticated mode. Nothing travels in clear between your teams and your agents.

At rest
Conversations kept by the platform are encrypted with AES-256-GCM, with a key stored outside the database.
Remote access
Through encrypted tunnels only. The gateway listens on the machine's local interface only.
Exposure
Any public exposure is prohibited in production.
Outbound flows
Encrypted and documented: the reasoning model, plus memory embedding in a standard installation. A sovereign installation blocks everything else.

The audit log is sealed.

Every action is timestamped, hashed with SHA‑512 and chained to the previous one with HMAC‑SHA‑512: any later change breaks the chain. The log can be verified from the cockpit and exported for your regulator.

Timestamped
Every decision and every action of the agent.
Chained
Each sealed event depends on the previous one.
Exportable
For your regulator, your auditor, your CISO.
Diagram: in an intact chain, each hash goes into the next one and the chain is valid. If an event is modified afterwards, its hash changes and the chain breaks from that point on.intact chainchain valid#8840collectionhmac 7b0c…#8841calculationhmac 3e1a…#8842reporthmac c95f…one event modified afterwardschain broken#8840collectionhmac 7b0c…#8841modifiedcalculationhmac 5d02…#8842reporthmac c95f…Diagram: in an intact chain, each hash goes into the next one and the chain is valid. If an event is modified afterwards, its hash changes and the chain breaks from that point on.intact#8840hmac 7b0c…#8841hmac 3e1a…#8842hmac c95f…chain validmodified#8840hmac 7b0c…#8841modifiedhmac 5d02…#8842hmac c95f…chain broken
Example.

No component raises its own rights.

The agent works inside a perimeter set for it. It can neither widen it nor touch the rules that bound it.

Read the details
Per-agent profiles
Allowed tools, safe binaries, restricted scope. A KYC agent does not have the tools of a reporting agent.
Locked configuration
Files locked at operating-system level: the agent can change neither its configuration nor its security rules.
Permission pipeline
Every sensitive action goes through fixed rules first, then a classifier for ambiguous cases. Destructive commands are refused and logged. When in doubt, the action is refused.

Who decides what.

  1. 1

    Your teamDecides what the agent is entrusted with. Can ask at any time for it to be stopped or replaced.

  2. 2

    Easylab AISets and deploys the rules, tools and permissions of each agent.

  3. 3

    The agentOperates within the set perimeter, with no autonomy beyond the permissions granted.

Hosted in Europe, with no public exposure.

Your agents run on dedicated machines, at Easylab AI or on your premises, in the European Union. Your data and your agents' memory stay in Europe.

Read the details
Provider
Easylab AI, Roeser, Luxembourg (RCS B276290).
Data and memory
Stay in Europe. Memory is stored locally, on each agent's machine.
Reasoning model
An interchangeable supplier, switchable to European-hosted models without re-architecting.

Easylab AI's list of sub-processors is published on the Compliance Hub.

The other layers.

What still protects, once everything else is in place.

Read the details
Deny list
Destructive commands are refused by fixed rules, built into the execution core, ahead of the classifier.
Blocking checkpoints
Every tool use is intercepted before execution, and can be blocked depending on context.
Locked-down workers
A worker whose authorisation checkpoint is missing does not start: never a permissive fallback.
Injection resistance
Multi-pattern filtering of external content, built into the core of orchestration.

For your file.

Download the System Card (PDF)