Security

Built for the firms that actually
read the security questionnaire.

When the data is bid-sensitive, LP-sensitive, or regulated, the security model is the product. This is how we think about it.

Core Principles

Four principles we hold
on every engagement.

No training on your data

Every model we use runs on commercial API terms with retention and training disabled wherever the provider allows it. Firm data never enters a training set. Period.

Data stays where it starts

Where possible, agents read from your existing systems (SharePoint, Egnyte, email, CRM) and write back to the same. We avoid creating new copies of sensitive data unless an engagement explicitly calls for it.

Least-privilege credentials

Every integration is scoped to the minimum read and write scopes required by the workflow. Credentials rotate on a fixed schedule. Access logs are on by default and shared with your security team.

Tenant isolation

Nothing about one client ever touches another. Separate credentials, separate compute, separate prompts, separate logs. No shared vector indexes.

Deployment Options

Where the pipeline runs is your call.

Engagements land in one of three deployment models. We recommend based on the sensitivity of the data and the tempo of the work. Every path is reviewed with your security team before the first line of code runs against real data.

ARC-hosted

We run everything. Fastest to stand up. Appropriate for public-market research, non-sensitive commercial real estate comps, and pre-production pilots.

Customer-hosted

Pipeline code runs inside your cloud tenant, your Entra, your subscription. Data never leaves your perimeter. Appropriate for LP, bid-sensitive, and regulated data.

Hybrid

Execution in your environment, orchestration in ours. Balances speed of iteration with information-security constraints on raw data.

Data Handling

Standards that apply to every pipeline we run.

Data minimization: we pull the minimum necessary to produce the work product and nothing else.
Prompt redaction: sensitive identifiers are tokenized before any model call and detokenized only on your infrastructure.
Output review: every production pipeline defaults to human-in-the-loop before deliverables leave the firm.
Incident response: a shared runbook is part of every engagement, with named points of contact on both sides.
What we are working on

Compliance roadmap.

ARC is an emerging firm. Formal certifications (SOC 2 Type I, then Type II) are on the roadmap as the client list grows. In the meantime, every engagement comes with a security addendum customized to your firm's requirements, a named point of contact, and a written data handling plan reviewed with your team before kickoff.

Need specifics for your security review? Contact us and we will walk your team through the architecture in detail.

Next Step

Pick a Workflow. Let's Build It.

A no-cost briefing is the start. Bring one deliverable that eats your team's time. We will sit with the analyst who owns it, map the pipeline end-to-end, and show you exactly what an ARC agent would do.