Every question we get in a security review

You define organisation-wide rules for models, sensitive data, agents and tools. obstruo enforces those rules on every request in the path, and writes the decision, the policy version and the routing into one evidence record.

No. We are neutral by design. We do not build models, we govern them, which is what lets you change provider without changing your controls.

Security, compliance and platform teams anywhere AI is in production. That includes organisations with regulated data, and any company serving AI in the EU, where the AI Act and GDPR apply whatever the sector.

Live: redaction, guardrails, models and routing, MCP governance, policies, audit, incidents and approvals. Private preview: agent registry. In development: human oversight, exposure map and kill switch. Planned: browser and coding-assistant adapters.

Most teams are live in a day: connect a provider key, point one application at obstruo, and inherit the organisation policy. Widening coverage is then a configuration change, not a project.

One base URL and one obstruo key. Existing OpenAI-compatible SDK calls keep working, and provider keys move out of your repositories into the vault.

A named, versioned set of rules: the redaction standard, the approved models, the permitted tools, the budgets and rate limits, and the behaviour when a check cannot be evaluated.

Organisation policies are marked locked, recommended or optional. Every project inherits them on creation. Locked policies cannot be edited inside a project, recommended ones can be declined with a reason, and optional ones are added by the project itself.

A project can add optional policies, decline a recommended one with a reason, and tighten anything further. It cannot loosen a locked policy. Whatever it changes is versioned and visible in the organisation view.

Roles are separated. Policy authors propose changes, compliance approves them, and every change is versioned with a diff and an actor. Admin activity is audited per action.

Each request is evaluated against the version active at that moment, and that version number is written into its record. Nothing is retroactively re-decided.

Through a dedicated request header that names the agent. The identity it points to carries an owner, a risk rating and a data classification, so policy applies per agent rather than per application.

Yes, by shaping the MCP server you hand it. From one connected MCP server you create narrower obstruo MCP servers that expose only selected tools, with restrictions applied, then point an agent at the one it should have. A read-only assistant might see 4 of 11 tools.

The call is denied before it reaches the tool, the attempt is recorded against the agent, and an incident can be raised automatically for review.

Yes. MCP servers and tool calls are governed alongside model calls, so the same policy covers what an agent can read and what it can do.

Spend ceilings and rate limits can be set per agent, per project or organisation-wide, with alerts before the ceiling and a configurable action at it.

Approval workflows for new models, new MCP servers and policy changes are live. Runtime human oversight of individual agent actions is in development, alongside the kill switch.

In front of your providers. Either your applications call obstruo directly, or obstruo is chained in front of the AI gateway you already operate.

Around 20ms p50 for the enforcement path. Redaction with entity detection adds more, depending on payload size, and is measured per project so you can see it.

You choose per policy: fail closed and refuse the request, or fail open and let it through. Fail closed is the default. Either outcome is recorded with the policy version that caused it.

Only in redacted form. Payloads are stored after redaction has run, so the entities your policy strips never land in our storage, and the record still shows what was sent, how it was handled and which controls applied.

Single-tenant and sovereign deployments are available on Enterprise, in an EU region of your choice or in your own environment.

Routing rules fall back to the next approved provider for that logical model name, inside the regions the policy allows. The failover is part of the request record.

OpenAI, Anthropic, Google, Mistral, Azure OpenAI and Amazon Bedrock, plus self-hosted OpenAI-compatible endpoints for models you run yourself.

Yes. You bring your own keys and commercial agreements, and you pay providers directly. obstruo governs the traffic and stays neutral about who serves it.

It can, but nothing forces you to swap. obstruo covers that job on its own if you want to consolidate, or it chains in front of the gateway you already run and handles policy and evidence while that keeps carrying traffic. Either way you can start with your higher-risk projects.

Today by export: pull the records per project, agent or framework and load them where your reviewers work. API access to the audit trail is planned, not live.

SSO through SAML and OIDC, with SCIM provisioning and role mapping on Enterprise plans.

Dedicated adapters are planned, not live. Anything that calls an OpenAI-compatible endpoint can already be pointed at obstruo and governed today.

Encrypted in a per-customer vault, never in your repositories. Keys can be rotated, revoked or swapped from the admin panel without a code change or a deploy.

Nobody, by default. Payload access for support requires your approval, is time-boxed, and is written to the same audit trail you read.

TLS 1.3 in transit and AES-256 at rest, with per-customer key separation for vaulted provider credentials.

Not yet. Dependency and infrastructure scanning runs continuously, and a third-party penetration test programme is being set up now. We will publish the scope and cadence once it is in place.

Write to office@obstruo.ai with "vulnerability" in the subject line. We work to coordinated disclosure and come back to you before doing anything public.

Role-based access, SSO enforcement, MFA for local accounts, session controls, and a per-action audit of administrative activity including policy edits and key rotations.

EU AI Act Article 14, GDPR Article 32, ISO 27001 A.8.16 and SOC 2 CC6.1. Records can be exported grouped by framework control.

Not yet. Neither certification is in place today, and both are planned over the coming months. We build and operate against those control sets in the meantime, and we will say plainly which controls are in place if you ask during a vendor review.

EU-hosted by default, in EU regions only. Traffic can be pinned to in-jurisdiction endpoints per project, and the processing region is part of every record. Hosting is EU-located with EU-resident storage, so the US CLOUD Act is not a route into your data through us.

Yes, with a current sub-processor list and standard contractual clauses where a transfer outside the EU is involved.

Every customer gets every template. GDPR, CCPA, LGPD and PIPL standards ship with the product, each following that jurisdiction's definition of personal data, and you choose which one applies to a project. Different projects can run different standards side by side.

Policy version, decision, routing, redaction counts, tool calls, incidents and approvals, linked per request and exportable per project, agent or framework.

Yes. One project, free, to evaluate the controls with real traffic. The full comparison lives on the pricing page.

Per plan, with limits on governed requests, projects and agents. Teams starts at €149 per month. Enterprise is custom, and covers organisation-wide and sovereign deployments.

No. You pay your providers directly for tokens. obstruo charges for the control layer, which keeps the incentive to route you to a cheaper model intact.

You are alerted before the ceiling. Enforcement never stops because of billing. Overage is either billed or capped, depending on the setting you choose.

Both. Annual billing gives two months free. All prices are in EUR and exclude VAT, which is reverse-charged for EU business customers with a valid VAT ID.

Upgrades take effect immediately through secure checkout and are prorated. Downgrades apply at the next renewal, so policy coverage never drops mid-period.