Data Path and Boundaries

AI Preview data path and boundaries

Concept #

Only administrator-triggered analysis of already ingested security logs is in scope. The provider sees redacted text if ai.redact stays enabled (usernames, private IPs, hostnames masked by default).

Manual analysis requires explicit acknowledgements: advisory_only_acknowledged, provider_retention_acknowledged, cost_bound_acknowledged, exact_ip_acknowledged as applicable.

Boundaries #

• No Hosting raw logs to the model
• No browser-held API keys
• No training export (training_export.enabled must be false)
• allow_private_provider_networks default false; requires offline_mode=false — high-risk override
• max_context_events default 100; analyze-incident 413 when over cap
• Circuit/rate limits per admin per minute
• enforcement_applied remains false on advisory results
• Internal AI development endpoints are not this data path

Common misconceptions #

• “Approval enforces.” It does not.
• “Offline mode means no network.” It means the configured provider must satisfy loopback validation unless you deliberately disable offline_mode.
• “Saving a provider starts analysis.” It does not.

Related #

• Limitations
• Status
• Operator control