Your telemetry is your data.
Production data is a meaningful trust decision. Here's how we handle yours.
Synthetic, public, or security-approved telemetry can be evaluated immediately. Production approval depends on your requirements. Google OAuth and scoped API keys are available; Microsoft login, SAML, SCIM, and a completed SOC 2 Type II report are not.
Review the controlled evaluation plan →What Anthropic receives, and how AI can be disabled
AI-written explanations use Anthropic Claude. Deterministic detection and RCA ranking run inside Epok. Here are the data boundaries and the controls available today.
Epok uses Anthropic Claude models for AI features: Haiku for routine drafting and Sonnet for Deep RCA and playbook import. Drafting can run automatically when incidents are analyzed, as well as when you request an AI feature.
RCA prompts include ranked causes, scores, service names, error categories, timing, metric and trace summaries, dependencies, and similar incidents with resolution notes. Deep RCA can also include your architecture context, matched playbook steps and links, RUM routes and impact, and affected customer names or IDs, tiers, and error counts when available.
Hypotheses list up to 5 ranked causes; Deep RCA lists up to 8. Cause titles and explanations are truncated after 150–200 characters, and up to 3 historical resolution notes after 200 characters. Truncation adds an ellipsis. Architecture context drops notes, then network details if it exceeds 8,000 characters; this is not a hard cap on the remaining context.
Log text can leave Epok. Deep RCA uses normalized messages from up to 15 alerts, each truncated after 200 characters. RCA can also include pre-error log messages truncated after 150 characters. Ask Epok result summaries can send up to 8 raw log messages, each limited to 240 characters, with timestamps, service names, and levels, plus your question and a result digest. Alert-title rewriting and noise scoring send normalized log patterns (up to 300 characters) and alert titles; pattern labelling sends normalized messages up to 500 characters. These are field limits, not a universal limit on the whole prompt.
There is no automatic redaction todayfor secrets or personal data before AI requests. Pattern normalization, truncation, and prompt-injection guards are not sensitive-data scrubbing. We do not control Anthropic's retention; see its commercial data-retention policy. Epok does not train models on your telemetry. Under Anthropic's commercial terms, API inputs are not used to train its models by default; see its policy pages.
There is no self-serve AI off switch. Operators can disable AI for an entire deployment. Workspace-specific disabling requires a deployment configuration change; it is not wired into the current deployment. If you need AI off for your workspace, contact [email protected] before sending data and wait for confirmation. Avoiding AI buttons does not stop background drafting. Deterministic detection and RCA ranking do not require an external model.
Encryption boundaries stated plainly
TLS in transit, modern cipher suites only. Sensitive configuration (notification webhooks, integration credentials) is encrypted at rest. Application database runs on a managed relational backend with provider-side encryption.
Tenant isolation on every query
Every read, every write, and every detection query carries account and project identifiers. Ingest streams are tenant-scoped. Concurrency limits apply per tenant so no one's workload starves another.
Retention and deletion
Retention is 14 days on the trial, 30 days on paid plans (Team and Growth). Ingest pauses when the trial ends; existing data remains readable while it ages out under the trial retention policy. Custom retention is reviewed per agreement. Tenant-scoped deletion is available on request.
Access controls
Login uses Google OAuth today. Microsoft login, SAML enterprise SSO, and SCIM are not available. API keys are scoped — read, ingest, or both — and listed per tenant. Alert state changes are timestamped with an actor; a full administrative audit log is not yet available.
Not yet available
- SAML enterprise SSO and SCIM provisioning.
- Microsoft login.
- A full administrative audit log; alert actions are recorded.
- A completed SOC 2 Type II report; work is in progress.
Compliance and deployment
SOC 2 Type II is in progress — we will publish the report here when it is complete. GDPR-aligned, DPA on request. Primary infrastructure runs in the EU (Germany/Finland). Custom residency, dedicated deployment, and self-hosting require an architecture and commercial review.
We publish our subprocessors
We publish the services involved in operating Epok and state the data each function receives. Most subprocessors receive account, billing, request, or product-usage metadata. When AI root-cause drafting is enabled, Anthropic can receive selected incident evidence, including representative telemetry messages, as described in the AI-handling section.
Have a security questionnaire?
We answer SIG, CAIQ, and custom questionnaires on request. Procurement, legal, and IT teams welcome.
Contact us →