AI agents operate where your SIEM has no visibility.
Your existing security controls were built for humans and predictable pipelines. AI agents read data, write code, chain actions across systems, and delegate to sub-agents — none of that shows up in your audit log. AgentLattice is the governance layer that closes that gap.
Security questions? security@agentlattice.com
Security Controls
What is in place today, not aspirational claims.
Infrastructure Security
- AES-256 encryption at rest (Postgres)
- TLS 1.2+ encryption in transit
- Logical tenant isolation via row-level security
- DDoS protection at edge
- Web Application Firewall (WAF)
- Supabase-hosted — no shared compute between tenants
Identity & Access
- MFA enforcement, admin-controlled
- Role-based access control (RBAC) with least-privilege defaults
- Scoped API keys with configurable expiration
- Keys hashed at rest (SHA-256), shown once at creation
- No standing production access for employees
- Session timeout and instant revocation
Audit & Monitoring
- Cryptographically chained audit logs — tamper-evident by design
- ECDSA-signed checkpoints for scalable verification
- Behavioral anomaly detection across 5 signal types
- Graduated enforcement: WARN → THROTTLE → HALT → KILL
- Separate tamper-proof enforcement log chain
- Webhook export for SIEM integration (Splunk, Datadog, Microsoft Sentinel, Elastic, Sumo Logic)
Data Handling
- AgentLattice processes action metadata only — never your agent's actual data
- What is stored: action type, resource category, record count, sensitivity level
- What is never stored: file contents, API payloads, database records
- Full subprocessor list published and updated
- GDPR-compliant — DPA template available for download
Operational Security
- 99.9% uptime SLA — documented RTO < 4h, RPO < 1h
- Documented incident response plan, 72-hour breach notification
- Background checks required for all employees with data access
- Annual security awareness training, tracked
- Managed endpoints (MDM) for all staff
- Third-party penetration testing in progress
SSO & Provisioning
- SSO via SAML 2.0 — on roadmap
- SSO via OIDC — on roadmap
- SCIM v2 provisioning/deprovisioning — on roadmap
- JIT provisioning via IdP assertion — follows SSO
- Contact us to discuss your IdP requirements before GA
Compliance & Certifications
Completed certifications and active programs in progress.
SOC 2 Type II report will be available to enterprise prospects under NDA upon completion. Request early access to readiness evidence.
Security Documentation
Pre-filled questionnaire responses and agreement templates to accelerate your security review.
Threat Coverage
The threats AgentLattice is purpose-built to address.
| Threat | Mitigation |
|---|---|
| Unauthorized agent actions | Every action evaluated against org policies. No policy match = action denied. |
| Privilege escalation via delegation | Monotonic scope narrowing enforced server-side. Children cannot exceed parent capabilities. |
| Audit trail tampering | SHA-256 hash chain with ECDSA checkpoint signing. Modifying any event breaks the chain detectably. |
| Compromised agent API key | Immediate revocation, 24-hour TTL on delegation tokens, cascade revocation for child agents. |
| Bulk data exfiltration | Behavioral baselines detect volume anomalies. Circuit breaker halts the agent automatically. |
| Prompt injection propagation | Cross-agent correlation detects anomalous behavior propagating through delegation chains. |
| Insider threat (human operator) | Segregation of duties enforcement. All human actions, including MCP operations, are audit-logged. |
| Agent self-replication | Hard limit of 10 active ephemeral children per parent, enforced at application and database level. |
AgentLattice governs agents that route actions through the SDK or MCP gateway. Agents that call external APIs without SDK instrumentation are out of scope. AgentLattice has the same trust posture as your identity provider — compromise of AgentLattice infrastructure itself is not within the threat model above.
Security reviews, pen test reports under NDA, vulnerability disclosure