Security in Artificial Intelligence Workflows and Agents

Looplex AI workflows combine documents, data, models, and tools to perform legal work. Security follows this process: it starts with information intake, continues through context selection and tool use, and extends to delivery of results and recordkeeping.

These controls apply to Copilot and Autopilot modes. In Copilot, the user leads the interaction; in Autopilot, the platform coordinates delegated work. In both, operations remain subject to the authorized scope and the organization’s policies.

External Content and Malicious Instructions

Documents, emails, web pages, and tool responses may contain instructions that attempt to alter the objective of a task or induce an improper action. This type of attack, known as prompt injection, can occur even when the user only requests an analysis of a file. OWASP describes this risk in applications that incorporate external content into model context.

Looplex identifies the origin of information and separates retrieved content from the instructions that define the execution. Inputs undergo checks and filtering appropriate to their format and the activity, including text extracted from images and documents.

Retrieved content does not grant permissions or, by itself, change enabled tools or required approvals. Authorization checks remain in the platform’s services, independently of the model’s interpretation.

These measures reduce the likelihood and impact of manipulation. Protection combines input, access, and execution controls because detection of malicious instructions may fail.

Authorized Context and Data Separation

Before making documents or excerpts available to the model, the platform verifies their association with the organization and the permissions applicable to the execution. Selection considers the purpose of the work and limits context to relevant information.

Separation covers repositories, search indexes, vector representations, caches, and execution state. Information available to one task is not automatically made available to another user, organization, or activity.

User collaboration memory follows its own access, write, and retention controls when the feature is enabled. External content is not automatically promoted to persistent instructions or user preferences.

The state of an Autopilot task preserves its progress. Resuming a task uses current permissions for new accesses and actions.

Tools and Execution Permissions

Each execution has an identity, a scope, and enabled tools. Integration credentials are managed by the responsible services, with restricted access and permissions appropriate to the purpose.

The platform validates the requested tool, its parameters, the resources involved, and authorization for the operation. Data produced by the model is treated as input that must be checked before reaching databases, external systems, or execution mechanisms.

MCP connectors and other integrations follow the same criteria. Tool descriptions, interfaces, and external components are assessed and versioned to prevent changes from inadvertently expanding an agent’s capabilities.

Network access and transmission destinations are limited according to the activity and configured policies. The presence of an address in a document does not authorize sending information to it.

Human Approval and Actions with External Effects

Workflows define review points according to the risk and nature of the action. Oversight covers operations such as sharing documents, modifying records, sending communications, or submitting materials for court filing.

An approval request presents the action, its scope, the recipients or systems involved, and the result that will be used. Approval is tied to these elements. Significant changes require renewed validation before execution.

In Autopilot, Workbench brings together monitoring, deliverables, and requests for human participation. In other workflows or integrations, review takes place in the environment designated for the activity. The organization defines responsible reviewers and approval authority, subject to product restrictions.

Execution Limits and Isolation

The platform controls time, resource consumption, number of calls, and retry attempts. These limits reduce the risk of endless loops, excessive consumption, and failures propagating between stages.

Activities that execute code or process files in working environments use isolation and restrictions on resources, network access, and credentials. Granted capabilities correspond to the task’s needs.

Before retrying an operation with external effects, the workflow checks the known state and the conditions needed to avoid duplication. An uncertain response from a connected system may require reconciliation or human intervention before another attempt.

Result Verification

Results undergo checks for format, consistency, and suitability for the next stage. Structured data is validated before being saved or used by other tools. Documents and analyses retain references to the sources used where applicable to the activity.

Legal review remains with the responsible professionals. Technical checks support that review, but the accuracy of a response, the suitability of a strategy, and authorization for an action require their own assessments.

Traceability and Incident Response

An execution records the components used, tool calls, state transitions, and relevant approvals. These records link a deliverable to the data, versions, and operations involved in the work.

Logging limits exposure of sensitive data and credentials. Access and retention follow the purpose of each category, distinguishing operational logs, security records, and documents or results that form part of the customer’s collection.

When suspicious activity is identified, Looplex can interrupt executions, restrict tools, revoke credentials, and preserve evidence. Suspending a task blocks further activity controlled by the platform; actions already completed in external systems are addressed through the applicable recovery or correction procedures.

Models and Providers

Routing uses the models and providers enabled for the product and deployment. Selection considers the purpose, data processing conditions, and policies applicable to the work.

Restrictions on models, location, retention, and external services are applied according to the contracted solution. The AI transparency note explains processing, training, and storage conditions.

Platform Standards and Responsibilities

Standard platform controls. Execution identification, permission enforcement, separation between organizations, tool validation, consumption controls, and event logging are part of workflow operations. Looplex administers these mechanisms and assesses changes to components under its management.

Configuration-dependent capabilities. Accessible collections, models, connectors, memory, permitted destinations, specific limits, and additional approval points vary by product, plan, and deployment. Configuration defines the scope of each task within the platform’s limits.

Customer responsibilities. The organization defines who may delegate, access information, and approve; administers authorizations in connected systems; and maintains appropriate playbooks and review criteria. Responsible professionals review results and actions, particularly those with legal, financial, or external effects.