Elyne Corp / agentic system security

EAS Defense

The agent proposes. EAS checks permissions, decides before an action takes effect, and keeps the evidence.

A local control plane for AI agents: identities, capabilities, signed policy, controlled execution, and a verifiable audit trail. Elyne Red Unified is included to challenge the system within an authorized scope.

Request an evaluation Explore the architecture

Deterministic decisionsSigned policyOS confinement tested in the labRed Unified included

Architecture

How EAS protects an agentic system

This animation shows five successive states: the agent alone, exposed flows, then EAS, Cyber, and OS confinement. Each stage lasts five seconds; the final view remains on screen.

Inputs

User prompt

Supplied or retrieved document

Another agent

API and MCP responses

EAS · control gateway

Identity verified · calls traced

OS confinement · managed process

AI agentreasons and proposes

Action checked before effect

Agent resources

MemoryContextRAGFiles

Reads and writes controlled

Outputs

Tools, APIs, MCP

Files and code

Decisions

Messages to agents

Action outside policy denied

▤ Signed evidence ledger

System monitored by Elyne CyberMITRE ATT&CK · system

EAS audits and decides.Actions are checked before they take effect; decisions and evidence are retained.MITRE ATLAS · attacks against AI

Elyne Cyber · complementary defense.A separate local service that remains active if EAS stops.

OS confinement in the lab.Signed Windows profile: one probe and its child verified. A complete agent remains to be qualified.

1/5 · Agent only

Controls

Authority separated from the agent

EAS connects decisions, execution, and evidence. Models may analyze or propose; the operator's policy defines what is allowed.

01 / POLICY

A verifiable rule before effect

Identities and capabilities are evaluated against a signed SecurityPolicy. The deterministic engine can allow, constrain, require review, or deny an action.

02 / BROKER

Tools under control

The ToolBroker checks tool requests and parameters. Connected sensitive actions pass through EAS authorization before execution.

03 / OS BOUNDARY

Native Windows confinement

For processes that EAS launches and manages, a signed profile can impose an AppContainer identity and limited OS permissions.

04 / EVIDENCE

An audit trail that can be checked

Decisions and events are stored in a chained ledger. Qualification correlates receipts, observed effects, and system state.

Deployment

Two integration levels

EAS can first mediate actions from an existing agent, then add an OS boundary when EAS launches and manages that agent's process.

Connected agent

The agent keeps its existing execution model. EAS decides and audits actions that pass through its interfaces; other paths must be connected or restricted by the environment.

Gradual integration of sensitive tools
Policy and evidence centralized in EAS

Managed process

EAS starts a process specified by a signed policy in a restricted Windows context. Lab work observed OS-level denials and checked the identity of a child process.

Permissions defined in advance
Target software qualified before deployment

Validation

Effects observed in the lab

Tests on a Windows clone correlate a request, EAS's decision, the observed denial, and the associated records. Evidence remains tied to the version tested.

File access

Direct denial observed

A confined probe attempted to access a protected canary. Windows denied the request and the canary's integrity was checked.

Network

Targeted outbound denial

A connection attempt from the confined process was denied, with a positive control to distinguish confinement from a network fault.

Child process

Identity retained

A child retained its AppContainer token, and its termination was checked when EAS stopped the managed process.

Operations

Services and ledger checked

After the tests, the lab checked the state of EAS and Elyne Cyber and verified the ledger.

Evidence scope: separate tests on the Windows clone, documented in qualifications R43, R44, R45, and R48. Full confinement of a third-party agent still requires qualification of its integration and permissions.

Adversarial validation

Elyne Red Unified comes with EAS

The defense is accompanied by a way to challenge it. Red runs authorized scenarios and records observations so a campaign can be replayed after each change. Elyne Lab Director can orchestrate and consolidate these evaluations.

Explore the three products

Scenarios aligned with MITRE ATT&CK and ATLAS
Real attacks within an authorized scope
Results and evidence kept distinct from test plans

Availability

Evaluate EAS in your environment

Windows R45 is available for a guided technical evaluation. The Linux port remains in laboratory qualification. Integration starts with the capabilities and resources that need protection.

Talk to Elyne Corp