SOC2 compliant. AES-256-GCM encrypted. Immutable audit trail. Because your AI agents handle your most sensitive data — and we protect it at every layer of the stack.
Application-layer AES-256-GCM encryption happens before any write hits the database. No cross-org key sharing. Ever.
Each organisation receives a dedicated AES-256-GCM encryption key. Breach of one org's data cannot expose another's — keys never co-mingle at rest or in transit.
AES-256-GCMOAuth tokens, TOTP secrets, agent memory tiers, execution logs, financial transaction data, API keys, webhook secrets — all encrypted before DB writes.
App-layer encryptionEncryption keys support rotation without service interruption. Re-encryption runs as a background job, old key versions are securely retired post-migration.
Zero-downtime rotationModern auth stack built on Better Auth with WorkOS SAML SSO. Humans and agents each carry distinct identity tiers with enforced data classification boundaries.
httpOnly cookie sessions (no localStorage tokens), TOTP-based 2FA with backup codes, API keys with per-key rate limiting and fine-grained scopes. JWT fallback for legacy integrations with automatic rotation.
Better AuthTOTPPlug-in enterprise identity: Okta, Azure AD, Google Workspace, and any SAML 2.0 provider. JIT provisioning, group-to-role mapping, and session duration controls enforced at the IdP layer.
SAML 2.0WorkOS4 human roles (Admin, Manager, Developer, Viewer) + 3 agent tiers (Tier 1: read-only reporting, Tier 2: operational transactions, Tier 3: strategic planning). Each tier enforces its own data classification ceiling.
Scoped API keys with expiry, per-key rate limits, and automatic revocation on suspicious patterns. All endpoints protected — no unauthenticated surfaces exposed, even internally.
Rate limitingScoped keysHash-chained, append-only audit log. Tampered records break the chain. Your auditors can verify every event without touching production systems.
Agent actions, human actions, financial transactions, security events, compliance checks, credential access, data exports — all captured with structured metadata, actor identity, and precise timestamps.
Each audit record includes the SHA-3 hash of the previous record. Any modification to historical records breaks the chain — detectable instantly. No silent data manipulation possible.
Structured exports for SOC2 assessors, formatted for common GRC platforms. Point-in-time snapshots, date range filtering, actor-scoped queries. Evidence generation takes minutes, not days.
The Medic agent runs nightly security patrols. Permission drift detection, expired credential revocation, audit chain integrity verification, anomaly flagging. Issues are auto-remediated or escalated to humans with full context.
A complete compliance framework — control matrix, risk register, incident response, disaster recovery, and continuous automated monitoring via the Auditor agent.
The Auditor agent runs nightly security patrols across your entire platform footprint. It checks permission drift, validates encryption at rest, verifies audit chain integrity, and flags deviations before your next assessment.
Expired credentials are auto-revoked. Misconfigurations are surfaced with remediation steps. Compliance posture is always current — not a snapshot.
Nightly patrolsAuto-remediationDefined RTO and RPO targets with documented runbooks. Automated backup verification, geo-redundant storage, and recovery testing on a regular cadence. Failover procedures tested — not just written.
RTO/RPO definedEvery third-party integration is assessed for security posture before onboarding. Ongoing vendor review cadence, sub-processor agreements, and data flow documentation for your DPA and privacy obligations.
Third-party riskAgents have real autonomy — but inside defined boundaries. Hard limits can never be overridden. Soft limits are configurable by org admins. Every decision is replayable.
| Type | Rule | Configurable? |
|---|---|---|
| Hard | No production data deletion | Never |
| Hard | No external client communications | Never |
| Hard | No security codebase changes | Never |
| Hard | No agent identity impersonation | Never |
| Soft | Max cost per bead / per run | Admin-configurable |
| Soft | Max consecutive autonomous beads | Admin-configurable |
| Soft | Max file changes per PR | Admin-configurable |
The enforceClientContentApproval middleware blocks all agent-generated client-facing content until a human approves it. No outbound AI content without a human checkpoint.
Step through any agent's decision tree post-hoc. See exactly what context it had, what tools it called, what it considered, and why it made each choice. Full interpretability for any action.
Decision auditAll data passing to AI model APIs is routed through a PII detection and scrubbing pipeline. Names, emails, phone numbers, SSNs, financial identifiers — replaced with tokenised placeholders before any LLM call.
Pre-LLM scrubZero-knowledge credential vault proxy architecture. Agents call the vault proxy — the vault calls the downstream service. Credentials stay isolated in a separate Bun service. Always.
The credential proxy runs as an isolated Bun service on port 8443, completely separate from the main application. It holds and uses credentials — agents only hold scoped, time-limited proxy tokens. Even if an agent is compromised, the credentials remain unreachable.
Emergency credential access with mandatory dual-approval, full audit trail, automatic expiry, and post-incident review workflow. Every break-glass event triggers immediate notification to security leads with complete context capture.
Break-glass usage is flagged in the compliance dashboard and included in the next audit report — no silent emergency access.
Dual-approvalFull auditEvery request traceable end-to-end. Betterstack structured logs, Sentry error tracking, Bun native WebSocket with org-scoped channels. Isolation at every layer.
Every log line carries a request-ID, org context, actor identity, and trace data. Correlated across services. Searchable, alertable, exportable. No needle-in-haystack debugging.
Request-ID correlationError events enriched with user context, org ID, request trace, and reproduction breadcrumbs. PII is scrubbed before transmission to Sentry. Alerting on error rate thresholds with Slack/PagerDuty integration.
PII-scrubbed payloadsMiddleware-enforced org-scoping on every query. WebSocket channels are org-scoped — no cross-org message leakage possible. API keys cannot access data across org boundaries even if compromised.
Middleware isolation
Request a security briefing, review the SOC2 documentation, or talk to our team about your specific compliance requirements.