AI Agent Security Whitepaper

AI Agent Security Whitepaper

Control what an agent can do

Control what an agent can do

An agent may be allowed to read a supplier’s report while the decision to accept the supplier’s risk stays with a person. That boundary has to survive tool calls, retries, and changes to the workflow. This whitepaper describes Zania’s security controls and the deployment practices to verify with your implementation team.

An agent may be allowed to read a supplier’s report while the decision to accept the supplier’s risk stays with a person. That boundary has to survive tool calls, retries, and changes to the workflow. This whitepaper describes Zania’s security controls and the deployment practices to verify with your implementation team.

AI governance

Classify AI use before choosing controls

Begin with who receives the output and whose authority the system uses. Individual AI, internal workflows, and production AI need different reviews. The same model or vendor can support all three, with different data access and consequences.

Individual AI: employee assistance

Individual AI supports one person, acts under their credentials, and returns results primarily to them. An analyst might use an assistant to summarize a supplier’s security report. Review the account and sign-in method, whether the report may be uploaded, and where its contents go. The analyst’s involvement does not make every action low risk: an assistant with permission to send messages or modify records needs boundaries around those actions.

Workflows: shared internal processes

Internal workflows automate a shared business process using a team or service identity. A supplier-review workflow might collect evidence, draft findings, and route them to reviewers. Its access can span several systems and continue after the person who started it leaves. Review the operating identity, permitted actions, approval points, and records needed to investigate or resume interrupted work.

Production AI: customer and partner services

Production AI powers a product or service for customers, partners, or the public. This includes releases being tested before launch. A customer-facing support agent needs review of the answers and actions customers experience, how its AI role is disclosed, and who handles failures. Evaluate expected and adverse outcomes before release, then monitor them in use.

These categories are independent of whether the system is built internally or purchased. They also differ from its deployment environment: a production use case may still be in a test environment. Review depth depends on its data, permissions, autonomy, and potential harm. Revisit the classification when its audience or authority changes.

Make the approval specific

An agent that drafts supplier findings for an analyst has a different scope from one that sends those findings externally or accepts a risk. Record the intended use, data, audience, allowed actions, and accountable owner. Decide who reviews consequential outputs and who handles disputed or incorrect results. Where people receive an automated message, make the agent’s role clear.

Review depth should reflect the possible harm and the reach of the system. Include privacy, fairness, and applicable legal requirements where the use case raises them. A security review cannot settle those questions through infrastructure controls alone.

A change can invalidate the original decision

Adding a write-enabled connector changes what a previously read-only agent can do. Expanding its audience or moving processing to another provider can change the review just as materially. Keep a record of model and tool providers, data destinations, and who must be contacted when a dependency changes or fails. Reassess the affected behavior before widening access.

Stanford’s AI Security Framework is a useful institutional example: it connects risk, deployment context, accountable ownership, and lifecycle review. Its deployment routes are specific to Stanford.

AI model processing and privacy

Start with what happens when information reaches the model. Where processing runs, how long context remains, and whether data is used for training are separate questions.

Private model hosting

Zania uses private, isolated infrastructure within Microsoft Azure for model hosting. Model processing is one part of the data path; the application also needs to control which documents enter a request and who can access the result.

No training on customer data

Zania uses customer data for inference: processing the information needed to complete a task. We do not use it to train, fine-tune, or improve foundation models. This commitment is separate from the retention of assessment records and logs, described below.

Request context

Request context is processed in memory and discarded when the request completes. This describes the inference context. Documents, assessment results, and application logs have their own lifecycle; they should be reviewed separately when agreeing retention requirements.

Customer separation and data access

A request should only use information from the customer workspace it belongs to. That boundary matters before retrieval, during processing, and when the result is stored.

Logical isolation

Zania uses a multi-tenant architecture with a dedicated workspace for each customer. Logical isolation separates customer data within that architecture. The relevant security boundary is the workspace: access to one customer’s documents must not grant access to another’s.

Least privilege

Internal agents and services receive the permissions needed for their assigned functions. Limiting those permissions reduces the data and operations exposed if a component is compromised. An agent’s ability to interpret a request should not determine its authority to act.

Data retention and deletion

An assessment may need to remain available for a later review. Temporary model context does not. Retention decisions need to account for both, along with the logs used to investigate activity.

Assessment data and logs

Zania supports lifecycle management for assessment data, with deletion available after an assessment completes and configurable retention for session logs. Agree the required retention period during your security review, including how deletion applies to uploaded documents, results, logs, and backups. Availability of deletion should not be read as automatic deletion of every record at task completion.

Model request context

Duration of request only

Post-assessment deletion

Available immediately

Session logs

Configurable retention period

Model training use

Never

Agent identity and authority

A person may be allowed to review a supplier without being allowed to approve it. An agent working for that person needs the same boundary. Trace each tool call to the identity that executes it, the permissions it uses, and the person responsible for the workflow.

Permissions that match the task

Zania uses role-based access control to assign permissions by responsibility. Assessment privileges can be scoped to the datasets, tools, and sessions needed for the work. For a supplier review, establish separately who can read evidence, edit findings, contact the supplier, and approve the result. A read-only task should remain read-only even when the agent proposes a useful next step.

Approval belongs to a specific action

In a deployment that permits external actions, an approval needs to describe what will happen. For an email, that means the recipient, message, and attachments. If those change after review, the earlier approval no longer describes the action being taken. Enforce the decision in the execution layer and retain the approved details. A model’s confidence in its answer is not authority to send it.

Keep delegation visible

An agent may act through a user’s delegated access or through a service identity. Record which applies, the resources and operations in scope, and how that access can be revoked. When one agent calls another, the receiving service still needs to enforce its own authorization checks. A handoff should preserve accountability without silently expanding permission.

OWASP’s agent security guidance describes scoped tool permissions and approvals tied to the action being executed. These are useful design criteria to verify for each integration.

Test behavior across the whole workflow

A plausible answer can conceal a failed control. The agent may cite the wrong report, read beyond its assigned workspace, or repeat a write after a timeout. Evaluation needs to inspect the records and actions produced by the workflow as well as the text it returns.

Treat retrieved content as evidence

A supplier report can contain instructions aimed at the model. Following them would let the subject of an assessment influence how the assessment is performed. Zania uses input sanitization and adversarial filtering to detect and block prompt injection. Permissions and action checks remain necessary when a filter misses an attempt. Material retrieved from a document should never acquire the authority of the user who requested the review.

Check the finding against its source

Zania’s generated responses pass through verification intended to identify harmful content, unsupported statements, and formatting errors. A reviewer still needs to inspect the evidence behind a consequential finding. A correct quotation from a report outside the period being assessed, or a report covering a different service, can support the wrong conclusion. Keep the source, its scope, and the reviewer’s correction together.

Enforce limits outside the model

For critical workflows, Zania uses deterministic code to validate AI decisions against predefined rules. The application decides whether a proposed operation satisfies those checks. Test that boundary with a permitted action and an attempt to exceed it. Check that an operation outside the workflow’s approved scope is denied, even when the model requests it confidently.

Build a test from a real assessment

Take a report that covers one service and ask the agent to assess a different one. Test separate variants with contradictory evidence, a missing attachment, and an instruction inside the document to suppress a finding. Each has an observable result: identify the service mismatch, surface the conflict, request the missing evidence, and refuse the embedded instruction.

Run the same assessment with a failed connector and an interrupted write. Check the downstream system before retrying. Record the tested configuration, observed failures, and release decision.

NIST’s AI Risk Management Framework treats evaluation and monitoring as continuing work across the system lifecycle. For a buying process, our AI GRC platform evaluation guide turns that principle into a practical vendor review.

Encryption and network security

Encryption protects data while it is stored and transmitted. Network controls limit the paths through which backend services can be reached. Both matter alongside identity and application permissions.

Encryption at rest and in transit

Customer data in Zania’s databases is encrypted at rest with AES-256. Connections between the client, Zania’s servers, and the Azure backend use TLS 1.2 or later to protect data in transit.

Encryption at rest

AES-256

Encryption in transit

TLS 1.2+

Cloud infrastructure

Microsoft Azure (isolated VNETs)

Backend connectivity

Private endpoints

Network boundaries

Zania uses Azure Virtual Networks and private endpoints to restrict access to backend services. Reducing public exposure narrows the paths an attacker can use. It does not remove the need to authenticate requests or enforce permissions within the application.

Evidence and accountability

A finding needs enough context for someone else to review it. An action needs a record of who authorized it and what actually happened. Decide which records the workflow must retain before an investigation makes the gaps apparent.

Audit logs

Zania records user logins, data uploads, and AI agent actions with timestamps and user identifiers. Immutable audit logs are available to customers for security reviews. During deployment, agree how these records connect to source evidence, approvals, and changes in the connected systems. The final answer alone may omit a failed call or an action whose outcome is still uncertain.

SOC 2 Type II report

Zania has a SOC 2 Type II report. Its scope, examination period, and findings help a buyer assess the controls covered by the examination. Review the intended AI workflow as well: a report does not establish that every new model, connector, or permission change behaves as expected. Request the report, network architecture documentation, and security policies through Zania’s Trust Center.

Stop safely and retire access

A stop control has a precise scope. It may prevent new work while a request already accepted by another service continues. Establish that behavior before an incident.

Contain the incident, then decide how to resume

Name the operator who can suspend the workflow or revoke its access. Test what happens to queued work, retries, and calls already in progress. If an agent sends the wrong supplier an attachment, stopping it can prevent another send; it cannot recall information the recipient has already received. Preserve the action record and follow the incident process for the disclosure.

Before resuming, establish which actions completed and which have uncertain outcomes. Check the connected system so a retry does not repeat a completed operation. Agree when work should move to a person and what evidence that person needs to continue.

NIST’s Generative AI Profile recommends rehearsed incident responsibilities and fallback arrangements, including dependencies on third-party AI services.

Retirement includes the access left behind

When a workflow is retired, disable its schedules, revoke its credentials and connector grants, and record who owns any unfinished work. Resolve the retention of uploaded evidence, derived results, and logs under the agreed policy. Removing the agent from a screen does not remove the permissions it held in another system.

SOC 2 Type II · AES-256 · TLS 1.2+