---
title: "AI governance and AI agent security | afka"
description: "How afka puts AI governance in code: sandboxed runs, scoped logins, approvals before anything irreversible, an audit log, and where SOC 2 stands."
url: "https://afka.ai/security"
lang: "en"
---

Security overview

# Put AI to work. Keep control of what happens next.

Give afka the context to help your business, with clear limits on what it can access, when a person must decide, and a record of the actions it takes.

[Start free](https://app.afka.ai/signup) [Book a demo](https://afka.ai/book-demo)

Protection in practice Example

The request

“Update the supplier’s bank details before the next payment.”

afka

## This change needs an owner or admin.

Changing where the company sends money waits for a separate decision.

Change identified

Decision assigned

Awaiting review · No change made

The person requesting the change cannot approve their own held step.

Access Action controls Audit trail Your data Assurance Documents

The right context. The right audience.

## Useful knowledge. Clear boundaries.

Connect what afka needs for the job. Company, group, and personal knowledge have different scopes, so sharing a workspace does not mean sharing everything.

01

### Keep each workspace separate

Database access rules separate records by workspace. Access is enforced below the conversation itself.

02

### Connect accounts deliberately

Company and personal connections have distinct permissions. Credentials are held separately from the model and added when the connected app is called.

03

### Choose what leaves the team

Guests on calls and visitors to your website have their own knowledge permissions. Internal knowledge is not automatically public.

Knowledge access Example

Personal knowledge stays with the member who added it.

Control before the action

## Keep routine work moving. Bring judgment to the exceptions.

Set each agent to Auto, Review, or Suggest. Sensitive actions still face additional checks, including when the agent is set to Auto.

Some decisions always need a person. Refunds above $100 per order and decisions against a candidate in hiring require approval in every mode.

A hold has a separate decision-maker. Payment destination changes, bulk deletion or export, and other potentially harmful steps can require an owner or admin.

Illustrative scenarios. Outcomes depend on the connected app, permissions, and the checks for the specific action.

Accountability you can follow

## See the action. See who was behind it.

The audit trail records the agent, the person it acted for, the action, its time, credit cost, and any linked approval.

Authorized members can export the log as CSV or NDJSON. An export includes all events in the selected period, including rows hidden from the on-screen view.

Review the security documentation

Activity record Example

09:41 · Requested by a member

### Payment details change requested

Agent: Operations · Company connection

09:41 · Action held

### Owner or admin decision required

Reason: payment destination change

09:45 · Decision recorded

### Request rejected

The payment details were not changed.

Append-only record CSV · NDJSON

Controls around the work

## Protection across the workflow.

### Protect sign-in

Members can use passkeys and authenticator-based two-factor sign-in with recovery codes. Each person can review and sign out their active devices.

Workspace-wide enforcement of two-factor sign-in is not available yet.

### Keep credentials out of prompts

Keys held by afka are stored in an encrypted vault. Connected-app credentials are applied at execution rather than exposed to the model.

### Separate code execution

Model-written code runs in an isolated environment for the task, with time, resource, and credit limits. The environment is destroyed after the run.

Network access is denied by default and restricted to the destinations needed for the task.

### Stop work when needed

Members can pause an agent. Pending approvals are cancelled, and remaining steps check the pause before continuing. Disconnecting an app revokes its connector grant.

Data protection

## Your data is here to do your work.

afka does not use customer data to train general-purpose AI models and does not sell it. The DPA also requires afka’s AI providers not to use customer personal data for general-purpose model training.

[Read the DPA](https://afka.ai/legal/dpa)

### Encrypted in transit and at rest

Traffic uses TLS. Encryption at rest is provided by the managed database, storage, and hosting providers.

### Storage and processing in the US

Customer personal data is stored and routinely processed in the United States. The primary database is in North Virginia. EU data residency is not offered.

### Retention and deletion

Meeting transcripts and speech samples have a 30-day retention period for both afka’s copy and the meeting vendor’s. Members can delete individual calls, tasks, and knowledge entries.

### Visibility into service providers

The Privacy Policy lists sub-processors and what they process. The DPA describes notice of changes and the available objection process.

[View the sub-processor list ↗](https://afka.ai/legal/privacy)

A clear basis for your review

## Know what is in place. Know what is still ahead.

afka does not currently hold a SOC 2 report or an ISO/IEC 27001 or ISO/IEC 42001 certificate.

Programme Status What this means today

### SOC 2 Type I & Type II

In progress

Work toward the examinations is under way. No report has been issued.

### ISO/IEC 27001

Planned

No certificate is held and the certification process has not begun.

### ISO/IEC 42001

Planned

The AI management system certification work has not begun.

**Additional details for your security review**

Independent penetration testing and a third-party security audit have not been completed. SCIM directory sync and workspace-wide enforcement of two-factor sign-in are not available today.

Audit exports are not cryptographically signed. The security policy documents additional operational gaps, including formal security programme and restoration testing coverage.

[Read the current control limitations ↗](https://afka.ai/legal/security#roadmap)

Take the detail with you

## Ready for a closer look.

Use the published documents for your internal review. Bring your security questionnaire or specific requirements to a conversation with our team.

### [Security policy](https://afka.ai/legal/security)

Technical and organisational measures, with current limitations.

Read the policy

### [Data processing agreement](https://afka.ai/legal/dpa)

Processing terms, responsibilities, transfer provisions, and Annex II.

Read the DPA

### [Privacy policy](https://afka.ai/legal/privacy)

Data handling, rights, and the published list of sub-processors.

Read the policy

Found a security issue? [Contact support@afka.ai](mailto:support@afka.ai) [Terms of service ↗](https://afka.ai/legal/terms)

Your questions, answered

## Before you connect your business.

**Do we need to connect every company tool?**

No. Connect the accounts needed for the work you want afka to do. Company and personal connections have separate permissions, and connecting one service does not automatically grant access to another.

**How do Auto, Review, and Suggest differ?**

Auto allows permitted external sends and writes to proceed. Review asks for approval, while Suggest presents a proposal. Reads, searches, drafts, and a reply the person requested can proceed in each mode. Mandatory holds and the checks for the specific action still apply.

**Can we delete our data or close a workspace?**

Members can delete tasks, individual meeting calls, knowledge entries, and their own account. An owner can close a workspace after signing in again. Closure includes a period for export or reopening before erasure; some records are retained in anonymised form. The Privacy Policy and DPA describe the applicable terms.

**Can we use afka for regulated data?**

The current DPA excludes specified categories of regulated data, including HIPAA-protected health information and payment card data. Review the DPA before connecting a workflow with sector-specific requirements.

**What can we review before getting started?**

The security policy, DPA with Annex II, Privacy Policy, and sub-processor list are published. Book a demo to discuss your requirements or request a completed security questionnaire. No SOC 2 report or ISO certificate is available today.

afka

## Bring your requirements. See how afka fits.

Walk through access, actions, and data handling with our team.

[Start free](https://app.afka.ai/signup) [Book a demo](https://afka.ai/book-demo)
