Security

Security that matches how Tilth actually works.

Tilth agents are designed to work inside clear boundaries: isolated client workspaces, human approval before external outreach, limited access to client systems, and an audit trail for the work they perform.

Talk to us about your setup Read the controls

The short version

Each client agent has its own workspace, memory, inbox, and portal scope.
Agents only receive the access needed for their assigned work.
External outreach is draft-first unless a client explicitly approves a different workflow.
Outbound messages pass through security checks before they leave.
Client portal data is scoped per client, with access controls at the database layer.
Tilth does not claim SOC 2, ISO 27001, or other certifications it has not completed.
"Can the agent send something without us approving it?"

Human approval comes first

For outreach and other high-trust workflows, Tilth agents prepare drafts for human review. The agent's role is to research, draft, organize, and surface the work. A person decides what gets sent.

For client-specific pilots, sending can be configured off entirely for the relevant workflow. In those cases, the agent produces reviewed deliverables inside the portal rather than transmitting messages itself.

Draft-first outreach Named human approvers Outbound message checks Approved-sender rules
"Can another client or agent see our information?"

Client data is separated

Each client agent runs with its own working space, long-term memory, email identity, and portal scope. Tilth does not use one shared client brain. Portal data is scoped per client, and database access rules enforce that separation on every query.

Separate workspaces Separate agent memory Separate email inboxes Client-scoped portal access Database row-level security No cross-client data pool
"How much access does the agent actually get?"

Least-privilege access

When an agent needs access to a client system, Tilth treats it like onboarding a new employee: named account, limited role, revocable access. We ask for the minimum access required for the work and keep the scope tied to the job the agent is doing.

Named accounts, never shared master credentials Role-limited access Client-controlled revocation Minimum necessary data No credentials over email or chat
"What if someone tries to trick the agent through an email or a web page?"

Outside content is information, not instructions

Tilth agents are trained to treat emails, web pages, attachments, and other outside content as information, not instructions. A page on the internet or a line in an email cannot override the agent's operating rules.

Untrusted-content handling Instruction-override refusal Suspicious input logging Output checks before messages leave
"Who is watching the work?"

Audit history and oversight

Tilth is a managed service, not unattended software. Client work is visible through the portal, including tasks, delivered files, chat history, and usage activity. For high-stakes work, a named human approver reviews outputs before they are used externally.

Portal task history Delivered-file history Chat history Usage activity Human review of high-stakes outputs Incident escalation
"What happens to our data if we leave?"

Retention, export, and deletion

Client data lives in the client's portal and agent workspace for as long as the work requires it. If a client offboards, Tilth exports the relevant work product and removes active client data from the system. Backup copies age out on the normal backup cycle.

Exact deletion timing and export procedure are confirmed in each client agreement before work begins.

The honest part

What we do not claim

Tilth does not claim to be unhackable, does not claim the AI is incapable of mistakes, and does not claim compliance certifications we have not earned. If your organization requires SOC 2, ISO 27001, HIPAA, or another formal framework, we will discuss that requirement directly before starting work.

Where our compliance work is headed

NowClear, published security controls and boundaries.
NextA stronger evidence trail: access, approvals, audit, retention, incident handling.
LaterSOC 2 Type 1, when client conversations make it worth the cost.
DeferredISO 27001, unless enterprise buyers specifically require it.

Have a security question before you start?

Tell us what your business needs protected, what systems the agent would touch, and who should approve external work. We will scope the setup before anything goes live.

Talk to Tilth