Skip to main content

AI Data & Privacy

Offload Security uses AI to help you work faster — summarizing scans, answering questions, drafting remediation and assessment content, and assisting triage. This page explains exactly what the AI sees, where your data goes, how it's isolated, and how to turn it off — so you can make an informed decision for regulated or data-residency-sensitive environments.

The short version

AI features are optional and assistive. They run only when your operator configures an LLM provider, they only ever see data your team already has access to, tenant data is never served from a shared cache, and remediation suggestions are advisory — a person makes the change. If you don't want prompts leaving your environment, you can run the platform with AI disabled.

Which features use AI​

FeatureWhat the AI does
AI assistant / chatAnswers questions about your security posture, using context from your own data.
Scan summariesTurns a completed scan into a plain-language summary.
AI SOC agentsAssist with triage, remediation guidance, compliance mapping and threat hunting.
AI-SPMJudges whether a prompt-injection test compromised a model (see AI-SPM).
Auto-fillSuggests answers for assessments and DPIAs, always for human review.
Knowledge Base Q&AAnswers from documents you upload (see Knowledge Base).

All of these are assistive. AI never changes your cloud, closes findings, or approves compliance on its own.

Model providers​

AI features call an external LLM provider's API. The platform supports:

  • Anthropic Claude (the default provider)
  • OpenAI
  • Google Gemini

Your operator chooses the provider and supplies the API key, which is stored encrypted at rest. The model is accessed over the network through the provider's API — the platform does not run a model locally.

What the AI sees, and where it goes​

  • The AI is given only data your active team already has access to — for example, the scan or finding you're asking about, or the documents you uploaded to the Knowledge Base. It is scoped to your team; it cannot see another tenant's data.
  • That context is sent to the configured LLM provider to generate the response, then returned to you.
  • How the provider handles data sent to its API is governed by your organization's agreement with that provider. Offload Security does not train models on your data; it uses the providers' inference APIs.

Caching and tenant isolation​

To cut cost and latency, the AI assistant uses a three-tier cache:

  1. Curated platform knowledge — generic, product-level answers shipped with the platform.
  2. Learned response cache — generic question/answer pairs learned over time.
  3. The LLM — used when neither cache has a match.

Tenant-derived content is never served from a shared cache — only generic, non-tenant answers are cached. Everything else the AI stores (conversations, summaries) is team-scoped and lives in your platform's own database, under the same tenant isolation as the rest of your data.

Turning AI off​

AI features are active only when an LLM provider is configured. If no provider/key is set, AI features are simply off and the rest of the platform works unchanged. To disable AI, remove (or don't configure) the provider key.

On-premises and data residency​

The whole platform can run fully on-premises — see On-Premises. One caveat to plan for:

AI features still call the external provider

Even in an on-premises deployment, AI features send prompts to the external LLM provider over the network using your key — the model itself is not run inside your environment. If sending prompts to an external AI provider is not acceptable for your data-residency requirements, disable AI features (above) and use the platform without them. Everything else — scanning, findings, risk, compliance, reporting — runs entirely within your environment.

Last updated on