AI assistant safety guide

Cyber Health Score includes AI-assisted experiences to explain the product and help workspace users interpret posture data. These assistants are designed as bounded application features, not as unrestricted access to your organisation's full backend or to other customers' data.

Two assistants, two scopes

Ask CHS appears on the public marketing site. It answers questions about Cyber Health Score capabilities, pricing, onboarding, and general defensive cybersecurity concepts using approved public product information. It does not have access to any customer workspace, tenant inventory, findings, scores, scans, evidence, or internal systems.

CHS Agent appears inside an authenticated workspace. It helps members understand workspace posture, findings, scans, compliance evidence summaries, notifications, and product usage within the permissions of the signed-in user. It retrieves and explains data through registered tools and approved summaries—not by browsing arbitrary database structures.

Some workspaces also use AI-assisted content generation for Authorised Phishing Simulation and Education (ASPE) email drafts. That flow is separate from the chat assistants: it runs only under explicit ASPE authorisation and workspace configuration, and it does not grant CHS Agent extra access beyond normal workspace rules.

Purpose

The assistants exist to:

• translate security and product concepts into plain language
• summarise posture and workflow context that the user is already entitled to see
• suggest practical next steps and in-product links

They are not autonomous security operators, penetration testers, or compliance certifiers.

Tenant-scoped context (CHS Agent)

CHS Agent operates in the context of the active workspace and the signed-in member's role (Owner, Admin, Analyst, or Viewer). Server-side checks resolve membership before any assistant tools run. The assistant cannot switch organisations or read another customer's workspace.

Ask CHS never receives tenant-scoped context from a signed-in session on the marketing site.

Approved data usage

CHS Agent may use tool outputs that are intentionally shaped for assistant use, such as:

• dashboard and posture summaries
• findings and scan context the user can already access
• compliance evidence summaries
• notification summaries
• glossary-aligned security term explanations

Tool results are passed through redaction and output-guard patterns so sensitive fields and forbidden keys are not exposed in replies. The model is instructed to stay honest when data is missing, stale, or a tool reports an error, and not to invent scan results or tenant-specific facts.

Ask CHS uses only public product knowledge and public tools appropriate to prospective visitors.

Permission boundaries

Assistant capabilities respect the same workspace authorisation model as the rest of the product:

• viewers see what viewers can see; admins see what admins can see
• the assistant cannot bypass role checks, support-access grants, or verification requirements
• the assistant cannot perform privileged actions on its own (for example changing billing, granting support access, or authorising scans) unless the product exposes that action through an explicit, user-driven workflow outside the chat

Recommendations from CHS Agent are advisory. Operational and security decisions remain with your organisation and qualified professionals.

Tool boundaries

CHS Agent interacts with backend data only through registered tools with defined inputs and outputs. Tools call tenant-scoped services; they do not expose raw internal schemas or arbitrary query surfaces to the model.

Ask CHS uses a separate, public tool set suitable for marketing-site questions.

Neither assistant is a general-purpose shell with unrestricted API access.

Human decision authority

Use assistant answers to inform prioritisation and communication, not as a substitute for:

• your own risk judgement
• legal or regulatory advice
• formal audit or certification opinions
• emergency incident response procedures

When the product offers approval workflows (for example operator-proposed actions), those workflows still require explicit human review according to product rules.

Auditability and rate limits

Assistant requests are subject to rate limiting at user and workspace levels to reduce abuse. Operational logging uses privacy-conscious patterns (for example hashing identifiers in logs rather than printing raw tenant identifiers in free text). This supports troubleshooting and security monitoring without treating chat content as a public log.

Failure and unavailable behaviour

Behaviour depends on configuration and environment:

• if AI services are not configured for an environment, assistant endpoints fail closed (typically unavailable responses rather than silent fallback to uncontrolled behaviour)
• when tools fail or data is not yet available, the assistant should say so clearly instead of guessing
• synthetic or demo fixtures, where present, are bounded and must not be mistaken for live model output in production customer workspaces

Availability of Ask CHS on the marketing site and CHS Agent in a workspace may differ by plan, feature flags, and deployment configuration.

Privacy and security considerations

• treat assistant conversations like other sensitive business communication; do not paste secrets, credentials, or personal data you would not store elsewhere
• Ask CHS must not be used to request another organisation's data or internal admin access
• CHS Agent inherits workspace isolation expectations described in Tenant isolation

What Cyber Health Score does not claim about AI

Cyber Health Score does not claim that its assistants:

• see everything stored in the platform database
• make autonomous security decisions or remediate systems without human involvement
• replace qualified security professionals or guarantee outcomes
• certify compliance or prove zero risk

For broader product boundaries, see What Cyber Health Score does not do.

Related documentation

• Tenant isolation
• Support access controls
• Roles and permissions