Security for
AI agent audit trails

Organisation and member records, billing information, admin activity logs, and agent audit entries: timestamp, agent ID, provider, model, token counts, tool names, approval decisions, risk flags, and SHA-256 hashes of prompts and completions. Oleus does not store raw prompts or completions. Your provider API keys are forwarded by the gateway and never stored.

If you route calls through the Oleus gateway, prompt and completion text pass through our servers in memory to compute hashes and PII flags, then are discarded. If you use the SDK wrappers instead, only hashes leave your process. PII flags are asserted by the SDK or gateway and are not independently verified.

In transit: TLS 1.2+. At rest: database and backups are encrypted by the hosting provider. TOTP secrets and SSO client secrets are additionally encrypted in the application with AES-256-GCM. API tokens are random 256-bit values.

Agent audit entries are append-only and kept for the life of the contract. Admin activity logs have a 365-day minimum, configurable up to 10 years. Deleted organisation data is purged within 30 days of account closure, except billing records kept for legal reasons.

Database hosting (managed PostgreSQL), AWS S3 (audit anchors), Stripe (payments), Postmark (transactional email), Vercel (website and dashboard hosting). Current list and regions on request.

Each entry stores the SHA-256 hash of the previous entry for that organisation, so any edit, deletion or reordering breaks the chain and is found by /agent-audit/verify. Database triggers block updates, deletes and truncation by the application role. A daily chain head is written to an S3 Object Lock bucket in compliance mode, so rewriting history would also require defeating that external copy.

Tamper-evident is not tamper-proof. A database owner can still alter rows; anchoring makes that detectable, not impossible. Truncation of the newest entries is only detectable after the next anchor. We do not claim immutability.

Only calls that go through the SDK wrappers or the gateway. Complete coverage requires blocking direct egress to model providers from your agents. Pass strict=True (or leave the gateway fail-closed) so an unrecordable call is blocked rather than skipped.

Exports are mapped to CC6.1 (access), CC7.2 (monitoring) and CC8.1 (change and approvals) to support your own audit. Oleus itself has not yet been audited; a Type I report is planned.

Yes. OIDC single sign-on (Okta, Google Workspace, Azure AD, generic OIDC) and SCIM 2.0 provisioning and deprovisioning.

Yes. TOTP for all users; admins can require it org-wide.

Yes. Configurable IPv4 addresses or CIDR ranges; other requests are rejected with 403.

Owner, admin, member and viewer, plus custom roles built from permissions: members:read/write, audit:read, billing:read and sso:manage.

Admins can issue per-agent API keys. The server fixes the agent ID from the key, records the key ID in every log entry, and ties each key to the member who created it; deprovisioning that member disables the key. Keys can be revoked immediately, and key creation and revocation are written to the hash chain. Stored as SHA-256 hashes; the secret is shown once. Without keys, the agent ID is a client-supplied label. Organisations can require keys.

Exports include columns mapping each entry to EU AI Act Art. 12 (record-keeping) and Art. 14 (human oversight), and DORA Art. 9 and 12, as evidence supporting your obligations. Oleus does not make a system compliant with either. Oleus stores hashes, not prompt content.

Sensitive tools (send_email, send_slack, refund, delete, update_billing, execute_sql, create_user, transfer_money) are held until a human approves in the dashboard. Approvals are bound to one agent and tool, single use, and expire after 15 minutes. The approver is derived from the authenticated session, not from client input.

Dependency scanning and CI checks are being rolled out as part of SOC 2 preparation. An independent penetration test has not yet been performed; we will publish its status here when it is.

Suspected incidents are escalated internally the same day and affected customers are notified without undue delay, targeting 72 hours or sooner where personal data is involved. Report vulnerabilities to security@oleus.io.

No uptime SLA is offered on standard plans. Contractual terms for Enterprise are discussed case by case.