Audit log
The audit log is an append-only record of decisions and privileged actions in Orchestrate. It lives in its own database, separate from run traces, so it stays when debug data is cleaned up.
Open Audit log in the console. It has four tabs: Events, Integrity, SIEM export and Retention. You need the platform admin role. In a multi-organization install, an organization admin sees their own organization's events.
How tampering shows
| Layer | What it catches |
|---|---|
| Hash chain | An event edited, removed or reordered. Each event's hash covers every field and the hash of the event before it. |
| Keyed hash (HMAC-SHA256) | A rewrite by someone with database access but not the chain key. A keyed event can't be rewritten as an unkeyed one either. |
| Signed checkpoints (Ed25519) | History rewritten after a checkpoint was signed, and the newest events deleted. The chain head is signed every 5 minutes when it has moved. |
| Append-only database role | The audit service itself can only read and add events, never change or delete them. Database triggers stop even the owner from changing an event, and allow deletes only below a retention checkpoint. |
| SIEM copies | A rewrite by someone who holds both the database and the keys: your SIEM still has the originals. |
Events written before the keyed chain existed keep their plain hash and still verify. The console marks them legacy.
Verify the log
- In the console: Integrity > Verify chain. The Events tab also checks the chain as it opens, and says whether it is intact or the sequence number where it breaks.
- From the command line:
voisctl audit verify --context <your-context>checks the stored chain inside the cluster. Add--checkpoint <file>to check it against checkpoints you exported earlier, from your SIEM or the console, and--jsonfor machine-readable output. It exits0when the chain is valid,1when it is broken, with the sequence and reason, and2on an error. - Through the API:
GET /audit-log/verify.
Integrity > Verification keys shows the public half of the signing key, so you can verify exported checkpoints anywhere.
No event lost
Every service writes each event to an outbox first, and removes it only once the audit service has accepted it. A relay resends with backoff, and the audit service drops repeats. So an audit service outage delays events but doesn't lose them.
Orchestrate also records an api.write event for every successful admin change that recorded no event of its own, and api.write.denied for every write refused for lack of a role. Data-plane calls, such as chat, the Agent API and channel webhooks, are left out of this sweep.
What is recorded
| Action | When | Decision |
|---|---|---|
chat, embed, transcribe, speak | Every model call's policy decision, with the model and, for external providers, the provider. Chat and speech calls also carry the PII types and counts found. | allow or deny |
guardrail_refuse | A PII guardrail refused a model call | deny |
guardrail_injection_block, guardrail_injection_flag | A call blocked, or flagged, as a prompt injection | deny, or none when flagged |
deploy_model | Every model deploy's policy decision | allow or deny |
model_image_attested, model_image_verified | A model build recorded its image digest; a deploy checked the image against it | |
agent_image_attested, agent_image_verified | The same for code agent images | |
call_tool | Every tool call's policy decision, and then the call itself | allow or deny, then none |
request_tool_approval, approve_tool_call, reject_tool_call | A tool call stopped for approval, and the decision | |
agent_call | One agent called another, over A2A or as a tool | |
eval_block_tool_call | An eval run skipped a write tool | deny |
agent.live_version.set, eval_gate_override, agent_settings.update | Release changes | allow |
custom_policy.* | Custom policies created, changed, deleted or imported | allow |
identity_provider.*, group.*, realm.security.update, tenant.mfa.update, account_link.* | Sign-in and access settings changed | |
auth.login, auth.login.failed, auth.logout, auth.mfa.configured, auth.mfa.removed, auth.password.changed, auth.password.reset_requested | Sign-ins and sign-outs, failed sign-ins, and MFA and password changes, copied from Keycloak | |
agent_channel.*, agent_source.*, code_agent.*, agent_secrets.update | Channels and code agents changed (secret names only, never values) | |
usage_budget.warning, usage_budget.refused | A budget passed its warning level, or refused a call | |
license.upload, compliance.export | A licence uploaded; an evidence bundle exported, with its digest | |
audit.export.configure, audit.retention.set | The audit log's own settings changed | |
realtime_session_event, realtime_tool_call | Events and tool calls of voice sessions and phone calls |
Each event has:
| Field | What it is |
|---|---|
sequence | Its position in the chain |
actor | Who did it: a user id, an agent key, or system for the platform itself |
action | What was done |
resource | What it was done to, such as the model, tool, project or agent |
decision | allow, deny, or empty |
metadata | Extra detail, such as a reason or the before and after of a setting |
tenant_id | The organization or project the event belongs to |
created_at | When it was written |
prev_hash, event_hash | The chain links |
Read the log
- Open Audit log > Events.
- Search by action, actor or tenant, and filter by decision. Open an event to see its resource and metadata.
The page shows the most recent events. To read further back, use GET /audit-log with filters for action, actor, project, decision and time range.
For a list of only what was refused, with a reason for each, open Governance and look at Blocked actions.
Export to a SIEM
SIEM export sends every event, and every signed checkpoint, to one or more sinks, in order and at least once. Each sink keeps its own place, and a new sink starts from the oldest stored event.
| Sink | Settings |
|---|---|
| Syslog over TLS (RFC 5424/5425) | Host, port (6514 by default), CA, format (CEF or JSON), facility |
| HTTPS webhook | URL, an optional bearer token, and an HMAC secret. Each delivery is signed in X-VoisX-Signature: sha256=<HMAC of timestamp.body>. |
| Splunk HEC | URL, token, index, source type and source |
| S3-compatible object storage | Endpoint, bucket, prefix, region and keys. Writes NDJSON. Optional Object Lock in governance or compliance mode, for a number of days. |
Add sinks in the console, and send a test event from each. Sinks your operator sets in the Helm values (audit.export.sinks) are shown read-only. Secrets are never shown again.
Retention
By default events are kept forever. A retention period can be set in the Helm values (audit.retention.days) and overridden on the Retention tab, but never below the minimum, 365 days by default.
A daily job verifies the chain first, deletes only events that every enabled SIEM sink already has, and writes a signed retention checkpoint that the chain then verifies from.
Keys
The chain key, the signing key and the write key are generated once on install and kept across upgrades, or you supply your own. To rotate the chain key, set a new one and list the old one as a previous key. To rotate the signing key, list the old public key. Dropping an old key makes every event or checkpoint it signed fail verification, by name. See Configuration.
What the log contains
Most events carry names, ids, decisions and counts. Text in event metadata, such as tool call arguments in approval events and transcribed turns in voice session events, is redacted for personal data before it is written, with the install's guardrail settings. Because the log is append-only, nothing in it can be removed later except by retention. Plan its storage and access with that in mind.
- A policy
denylooks the same whether a rule refused the call or the policy engine couldn't be reached. - The chain is one sequence for the whole install, not one per project.