Security at Myrmex
Giving AI the ability to act requires clarity about access, credentials and accountability. See how Myrmex organizes these boundaries, from the connected environment to the change record.
Credentials managed by integrations are resolved server-side. The AI agent references the integration by ID without receiving the secret.
How integrations workAI guardrails
Before acting, Myrmex considers context, impact and reversibility. Guardrails combine permissions, risk assessment, human approval and execution containment.
Illustrative example
The query can proceed with the required permissions. Reading an asset does not grant permission to modify it.
Queries follow granted permissions. Write actions require a plan and human approval; irreversible operations with maximum impact are blocked by the engine.
The kill switch contains autonomous execution and prevents new actions from being dispatched. Time and inactivity limits can also terminate execution.
Stopping does not undo a change already completed on the target. Reversal requires a compensating action with its own approval and record.
Interaction, decision and execution records support investigation. Operator feedback and reporting channels help investigate unexpected behavior.
Role-based access separates reading, execution and administration. Asset groups define accessible resources. In Enterprise, attribute policies add conditions to the permissions granted by a role.
Enterprise identity controls
Connect authentication to your corporate identity provider.
Define permitted actions by role and the accessible asset groups.
Add conditions such as source network, time, MFA and device.
In change management, a plan goes through approval before execution. The context’s rule defines who approves and how many approvals are required. Explore what happens in each situation.
Standard changes require execution to be started after approval. Approved patches may install automatically in their scheduled maintenance window.
An execution may finish as completed, partial or failed. Approvals and rejections record the user and time to support review.
Illustrative example
Context: production
Example rule: two named approvers
The change is not yet ready to execute.
Security owner: decision pending
Sample data. Separation of duties depends on configured permissions and rules. The default allows one approval, including by the requester.
A shared service account and an individual credential serve different needs. For compatible integrations, choose the model that fits your operation.
This is the default model. People with integration access use the credential configured by the organization. That account’s privileges need to match the operational scope.
Read documentationThe integration service resolves the credential to execute the action. This protection applies to managed credentials; documents, tool outputs or secrets typed into a conversation require care with the content you share.
Private mode depends on connector support and a valid credential linked by each user.
Cryptography and evolution
Data protected in transit and at rest, with an architectural direction guided by crypto agility: adapting cryptographic mechanisms as risks and standards evolve.
Protection in operation
Transport encryption, stored-data protection and device identity address different parts of the architecture. Keys, algorithms and protocols need to be assessed in context.
Architectural direction
An architecture guided by cryptographic agility prepares for changes to algorithms, protocols and key management. Each transition considers compatibility, dependencies and operational continuity.
Crypto agility also guides preparation for new standards, including post-quantum cryptography. Adoption is assessed per component against deployment requirements.
Reference: NIST crypto agilityThe agent installed on the host establishes an encrypted outbound connection to the platform. A newly enrolled device needs manual approval before becoming active.
Approved device
Telemetry and commands on the established channel
The key pair is generated on the host. The private key stays on the device.
A session key protects the data sent by the agent.
Traffic travels through an encrypted transport channel to the platform.
The Collector reaches internal systems over SSH or APIs. Their access permissions and firewall, proxy and TLS inspection rules need to be part of the deployment design.
Myrmex keeps operational context segregated by organization. Conversations and reports can form this memory, helping teams resume investigations and decisions with continuity.
Data, availability and support requirements are addressed during the review. This page does not replace contractual commitments.
Retaining context to support your operations is different from training model weights. During an Enterprise assessment, we formalize non-training commitments, data handling by AI components and the responsibilities of the services involved.
Processing and storage locations, retention, deletion and support access should be assessed for your scope. Specific requirements should be included in the proposal and contractual documents.
Match operational history to your audit, retention and incident response requirements. Centralized audit logs are an Enterprise capability; immutability requirements and export formats need to be confirmed.
Bring a use case and your requirements. Start with documented mechanisms, then work through the specific conditions your organization needs.
Scope, roles, attributes and authorization criteria.
Context-specific rules and recorded change decisions.
Enrollment, cryptography and outbound connection.
Shared accounts, private accounts and secret resolution.
Security scans in the CI/CD process before updates are released.
Digitally signed binary, hash verification and progressive update rollout.
Internal and independent testing covers the product and AI components. Evidence can be requested through a confidential channel.
Auditing framework controls or producing evidence does not amount to certification of the product or organization.
Queries can proceed within granted permissions. Write actions require a plan and human approval. Change management can add approvers and rules: standard changes are explicitly started after approval; previously approved patches may run during the scheduled maintenance window.
Standard SSO and RBAC are included in Business and Enterprise. Attribute-based policies (ABAC) belong to Enterprise scope. Every plan includes permissions and change authorization. The design also considers asset groups, workspaces and privileges granted in the source system.
For agent-to-platform communication, the host initiates an outbound connection on TCP 443. The established channel carries telemetry and commands. A Collector deployment also needs access to internal systems over their respective protocols.
Review the scope, permissions, proposed plan and approvals before execution. Start in a bounded environment and verify the result of each change. Controls support that supervision; they do not guarantee that every response or proposed action will be correct.
From your first connected environment to Enterprise requirements, let’s examine the scope with your team.
Myrmex is developed by AINEXT Tecnologia. Meet AINEXT
