Shared responsibility
Elitea enforces controls at the points it owns: model traffic, tool calls, credentials, runs and configuration. It records what happened as evidence, and exports that evidence. It does not try to replace the systems a regulated organisation already runs for reviews, records, identity governance or security monitoring. Rebuilding those inside an AI platform would produce a weaker copy that your auditors would then have to assess separately.
The table shows where the line sits and how Elitea integrates across it.
| Responsibility | Lives in | What Elitea provides |
|---|---|---|
| Architecture, legal and compliance, and model-risk reviews | Your GRC or ticketing system | Typed attestations on the agent or release, required by risk tier before approval, and writable over the API (attestations) |
| Data owner approval and data accuracy | The data owner and your data catalogue | Credential use grants with an approver other than the proposer (credential governance); data_owner and data_accuracy attestations with an expiry (attestations) |
| Immutable storage of record (WORM) | Object storage with object lock, or your SIEM’s archive tier | Hash-chained logs that make tampering detectable, exported in one versioned schema (audit event schema). Elitea does not claim WORM on its own database. |
| Retention schedule | Your records-management policy | Per-class retention limits and legal hold that enforce the values you set (audit evidence and retention) |
| Security monitoring, correlation and anomaly detection | Your SIEM or SOC | The evidence export (including every agent run’s step records), Prometheus metrics, alert rules for platform health, a per-agent Grafana dashboard, and a rule-based agent circuit breaker that stops a failing agent. |
| Paging and on-call | Your incident-management tool | Alerts (Prometheus rules for breaker trips, breaker storms and active suspensions; a platform stop is critical), the automatic stop of the agent circuit breaker, and the emergency stop that on-call staff use. Elitea does not page a person: route the alerts to your on-call. |
| Rotation of externally issued secrets | Your secret manager (Vault or OpenBao) or the issuing system | The secret-store integration, expiry on tokens and grants, and revocation (secrets) |
| Content classifiers and harm taxonomy | Your trust-and-safety or compliance function | The gateway enforcement point, built-in regex and secrets detectors, and the HTTP adapter to the classifier you choose (Guardrail pipeline) |
| AI disclosure wording per jurisdiction and contract | Your legal function | The disclosure marker on outbound text, with the text you set, a project override that cannot remove it, and opt-outs that only a platform administrator allows |
| Organisation-wide rules for agents | Your policy owners | The platform preamble before every agent’s system prompt, versioned with an audit event per change |
| Content classifiers and harm taxonomy | Your trust-and-safety or compliance function | The gateway enforcement point and adapters to the classifier you choose (planned, issue 1500) |
| Statistical drift on production traffic | Your ML-monitoring tool | The evidence stream as input, and an API to mark an agent as due for recertification (planned) |
| Bias and fairness methodology | Your model-risk function | An evaluation harness that runs any suite and scoring dimension you define |
| Business thresholds (amounts, novelty, risk appetite) | Your risk owners’ written policy | Enforcement of the thresholds and tool categories you configure |
| Sign-in frequency and multi-factor authentication | Your identity provider | Session maximums and an audit event for every session start, expiry, sign-out and revocation (session policy) |
| Access reviews and entitlement campaigns | Your identity governance (IGA) tool | The entitlement export as CSV and SCIM, the credential inventory and the permission trimming report |
| Architecture review of new connectors | Your architecture or security review board | Connector approval with a second approver and an optional mandatory attestation |
| Network perimeter | Your network team | Egress guard, the worker egress allowlist and the record of hosts each run reached |
| Trust in remote agents and their issuers | Your security architecture and the owners of the remote systems | Signed assertions on calls to remote MCP servers, a list of trusted issuers that an administrator keeps, and refusal of replayed or expired assertions. You decide which issuers to trust and verify assertions on your own servers. |