Skip to Content

Security and data handling

For a security review: how Agent Authority encrypts credentials, isolates organizations, and keeps an audit trail

This page lists how Agent Authority protects stored credentials, keeps each organization's data separate, and records and redacts what agents do. Use it for a security review. For certification status (SOC 2 and similar), contact your SecureAuth account team.

Organization isolation

Agent Authority is multi-tenant: every request belongs to exactly one organization. Gateway requests take it from the address they arrive on, where your organization's slug is the leading label of the service domain. Console and API requests take it from the signed-in user's token. SecureAuth scopes stored data to that organization. Credential encryption at rest enforces the same boundary a second time, so even a leaked ciphertext won't decrypt in another organization's context.

Credential encryption at rest

SecureAuth encrypts resource credentials at rest with AES-256-GCM. These are the OAuth tokens and any other secrets stored for a connection. Each ciphertext is bound to its organization through authenticated associated data (AAD). A value that decrypts in one organization's context can't be replayed in another's.

The encryption layer supports key rotation. It encrypts new values with the current key, and tries each previous key in turn when decrypting. This lets SecureAuth rotate a key without re-encrypting or invalidating credentials written under an old one.

Encryption in transit

The gateway serves traffic over TLS.

Per-user credentials, not shared accounts

The gateway calls each catalog resource that needs a sign-in, such as Slack, GitHub, or Jira, with the requesting user's own connection, never a pooled or shared service account. It acts on behalf of each user. Each user reaches only what their own account can see. The upstream service attributes each action to the right person, and revoking one user's connection affects only that user. Resources that need no sign-in, such as Microsoft Learn, get no credentials. See Per-user isolation.

The exception is custom resources configured with API Key or OAuth2 client-credentials authentication: those credentials are set once on the resource and used for every user who calls it, so upstream attribution is per-resource, not per-user. (A custom resource using the OAuth2 authorization-code grant is per-user, like the catalog.) A custom resource with no authentication sends no credentials. The audit log still records which user and agent made each call.

After a user connects a resource, the gateway sends them back only to an https address on its own service domain or a configured allow-list. It rejects protocol-relative addresses.

What the audit log stores, and for how long

The audit log is the audit trail of every gateway call. Each entry is structured event metadata: a timestamp, the actor (user and agent instance), the action type, the resource involved, and action-specific details such as tool name or policy effect. Entries made from a request also carry its context, when available: IP address, approximate location (city and country), device, OS, browser, user agent, and a trace ID.

By default it stores only the record of the gateway call itself: who made it, what it targeted, and what the policy engine decided. It stores the prompt of a model call only when an Audit Policy covers that call, and never stores the response text.

SecureAuth keeps audit events for 3 months, then deletes them automatically. LLM telemetry follows the same 3-month retention. To keep evidence past that window, forward events to your own system as they happen with webhooks.

Response redaction (DLP)

Data Protection response filters rewrite sensitive text (SSNs, API keys, email addresses) out of a tool's response before it reaches the agent. Redaction runs inside the gateway, so the sensitive values never leave it. The audit record of a redacted call stores the rule names that matched and how many spans each replaced, never the matched or redacted text itself. If a filter's pattern can't be compiled, the gateway blocks the call rather than forwarding an unredacted response. The one exception is a rule that also has a judge filter: it follows that rule's If the judge fails setting.

A judge filter can't run when no model is available to it, for example when no candidate is active. It then follows its rule's If the judge fails setting. The default blocks the response. See Failure handling.

Gateway addresses by region

The gateway itself is served from a region-specific service domain, with your organization's slug as the leading host label; the OAuth callback substitutes oauth for that label. Gateway-to-resource traffic goes to each provider's own endpoint, but some providers require allow-listing the gateway's callback origin on their side. Atlassian, for example, requires allow-listing https://oauth.aisecurity.services.<region>.connect.secureauth.com/**. See Atlassian's dynamic client registration setup for your region's value and the full steps.

Next steps

On this page