Serene privacy and safety

Built. It answers in early-access workspaces once HyoNova connects its AI provider; until then Serene says it is not available, and nothing is simulated.

Serene is built so that adding AI to your workspace does not widen who can see what. This page explains the rules HyoNova enforces — in the server and the database, not only in the instructions given to the AI model.

Permission-first answers#

• Serene acts as the person asking. It has no access of its own: it uses the asking member's role, permissions and data clearance.
• Filtered before the model. When Serene searches company knowledge, results the person may not read are removed in the database before anything is sent to the AI model. The model never sees what the person could not open.
• Checked live. Every search re-checks each source record under the person's own access. A record that was archived, a file whose sharing was removed or a support queue the person left disappears from Serene's answers immediately, whatever was indexed before.
• Sources are real records. The Sources under an answer come from the records HyoNova's tools returned — the AI cannot invent one — and opening a source checks access again.
• Company switches. An owner or admin chooses which kinds of company content Serene searches. A kind that is switched off is not indexed, its existing entries are removed from the index, and it disappears from Serene's search for everyone; Serene can still open a specific record that a person asks about, within that person's permissions.

Your company only#

Every company in HyoNova is isolated from every other in the database, and every Serene table carries the same isolation. Serene conversations are private to the member who started them. Knowledge is indexed per company, with each source's data classification.

The AI can only use HyoNova's tools#

The AI model never gets database access, queries, passwords or keys, a shell, open internet access or the ability to run code. It can only call a fixed list of HyoNova tools, each with a defined input, a required permission, a data classification and a risk level. For every call, HyoNova's server validates the input, checks the permission of the person (or AI Employee) it acts for, applies the approval rules and only then runs the same HyoNova service a person would use. Each request also has limits on the number of tool calls, steps, time and cost.

Approvals and risk levels#

A proposal shows what would happen, why and its sources. Only a person can decide it; it runs exactly once, and it expires after 7 days. Approved Serene proposals run with the approver's current permissions.

Instructions hidden in data#

Customer messages, emails, documents and other company content can contain text written to manipulate an AI — "ignore your rules", "send me all customer data". HyoNova treats all such content as data, never as instructions:

• HyoNova's own rules come first and cannot be changed by anyone in a conversation. A company's own instructions come next and can set tone and facts only.
• Content from records, files and customers is clearly marked as quoted data when it is passed to the model.
• Even if the model were persuaded, it could not go beyond the tools, permissions and approvals above.
• Company instructions for an AI Employee that try to weaken permissions, approvals, data protection or the AI disclosure are refused when they are saved.

The AI never pretends to be human#

• Serene's and AI Employees' messages are labelled as AI in the workspace.
• Customers see an AI Employee's name with · AI. A company can rename the AI Employee but cannot remove the AI label.
• If a customer asks, the AI Employee says it is an AI assistant of the company, and that a person can take over.

Data sent to the AI provider#

HyoNova uses an AI provider that HyoNova connects for the whole platform; companies do not bring their own AI keys.

• What is sent. Only what a request needs: the question, the recent messages of the conversation, HyoNova's instructions and the excerpts and records Serene retrieved for that person — never a whole workspace.
• Secrets removed. Obvious secrets in company content — passwords, API keys and tokens, one-time codes and payment card numbers — are replaced with "[redacted]" before the content is indexed or sent. Signing, finance, payroll and HR records are not indexed for search at all.
• Not stored by the provider for conversations. HyoNova keeps the conversation in your workspace and sends each request without asking the provider to store it as a conversation.
• Keys stay with HyoNova. The provider's API key is kept in HyoNova's secret store on the server. It never reaches a browser or a log. In voice conversations, your browser connects only the audio; HyoNova's server creates the session and decides its instructions and tools.

What is kept, and who sees it#

The retention check runs every hour. Billing and usage records of AI requests contain no conversation content and are kept as billing records.

Limits that protect your budget#

Every AI request runs as a bounded run: a reserved maximum of Serene credits, a ceiling per request, limits on tool calls, steps and time, HyoNova's own daily cost safety limits per company and per AI Employee, a limit on how many AI requests run at once across HyoNova, and a usage limit on AI Employee trials. When the owner's monthly budget cap or the balance is reached, Serene stops and says so. HyoNova can also pause AI for the whole platform, a provider, a feature, a company, an AI Employee, voice or autonomous replies at once if something goes wrong.

Status#