At a glance
Pintle is built for buyers whose security teams pre-approve every vendor. We’ve designed the platform — and our company — for the questions you’re already asking.
| Area | Posture |
|---|---|
| SOC 2 Type II | In progress; annual audit cycle |
| ISO/IEC 27001 | In progress |
| NIST AI RMF | Controls mapped (Govern, Map, Measure, Manage) |
| EU AI Act | Risk classification & obligations documented |
| Data residency | EU and US regions; single-tenant available |
| Encryption | TLS 1.3 in transit; AES-256 at rest; KMS-managed keys |
Status reflects current posture; updated each quarter. Attestation reports and security questionnaires are available under NDA — contact us.
Data handling
- No training on customer data. Customer prompts, completions, and tool inputs are never used to train Pintle’s models or any third-party model we route to.
- Configurable retention. Hot retention defaults to 30 days; cold retention defaults to 7 years for regulated industries. Both are configurable per workspace.
- Sub-processor transparency. A live list of every sub-processor (model providers, infrastructure, observability) is published and version-controlled. You’re notified before we add a new one.
- Regional pinning. Customer data stays in the region you select. Cross-region replication is opt-in.
Model governance
Pintle Control is the single chokepoint for every model call made on the platform. That means:
- Per-team allow-lists. Each team can use a specific set of models; nothing else routes.
- Per-data-class rules. Sensitive data classes (PII, PCI, PHI) can only flow to models that meet the corresponding compliance posture.
- Evals as guardrails. You can require evaluation pass rates before a flow is promoted to production.
- Drift detection. Production traffic is sampled and re-scored continuously; alerts fire when scores regress.
Audit trail
Pintle Insights records every prompt, tool call, decision, and policy evaluation as a signed event in an append-only ledger. The ledger is:
- Tamper-evident. SHA-256 chaining detects any modification.
- Identity-attributed. Every event carries the verified identity of the user, agent, or service account that initiated it.
- Exportable. Stream into your existing SIEM (Splunk, Datadog) or warehouse (Snowflake, BigQuery, S3).
- Framework-mapped. Pre-built views align to SOC 2, ISO 27001, NIST AI RMF, and EU AI Act control families.
Identity & access
- Federated SSO. Okta, Azure AD, Google Workspace, plus any standard OIDC or SAML provider.
- Scoped service identities. Each agent runs under a service identity with narrowly scoped permissions and short-lived tokens (default 15 min).
- Least-privilege defaults. New workspaces start locked down; access is granted explicitly.
- Break-glass workflow. Emergency elevation paths are themselves audited, time-bound, and reviewed.
Responsible disclosure
If you’ve found a vulnerability, please email us at security@pintle.ai. We respond within one business day and aim for a fix or mitigation within seven days for high-severity issues. We don’t yet operate a public bug bounty; please don’t publish before we’ve had time to remediate.
Talk to us
Detailed control mappings, penetration test summaries, and security questionnaires are available under NDA. Get in touch.