Security
How Levirge services are built and operated: workspace isolation, token handling, encryption in transit, source-protecting limits — and the risks intrinsic to LLM agents, stated plainly.
Updated 2026-08-08. This page describes what is in place today, in plain language. We do not hold SOC 2, ISO 27001, or comparable third-party certifications — the sections below say what we actually do.
At a glance
Levirge services (Search, Brain) are hosted services we operate. Workspaces are isolated per customer, access is by per-user tokens you issue and revoke, all public endpoints run TLS, and every read a Search workspace performs is recorded — URL, outcome, timestamp; the page as the agent received it stays available for 24 hours, then expires with the cache.
Workspace isolation
Isolation is enforced by the platform, not left to application code. A Search workspace's cache and stored results are per-tenant — there is no shared cross-tenant cache. Brain knowledge is partitioned into Vaults scoped to a tenant, with personal Vaults private to their owner and enforced fail-closed.
Authentication and tokens
Sign-in to the consoles is via hosted identity providers. Machine access is by opaque bearer tokens issued per user; only the hash of a token is stored, never the raw credential. You issue and revoke tokens for your own workspace, and usage is visible by user.
Encryption in transit
All public endpoints serve TLS with certificates issued by Let's Encrypt; deploys are verified against the public endpoint, including the certificate.
Limits that protect sources — and you
Search reads public pages a person could open in their own browser. Requests are paced per source with jitter at a human pace behind a concurrency guard, and read content expires from the cache after 24 hours. These limits are part of the service, not options.
Risks intrinsic to LLM agents
Content fetched from the open web is untrusted input. A page can contain text addressed to your agent — instructions, links, requests. Search returns content faithfully and records what was returned; it cannot make a page trustworthy. Treat fetched content as data, not instructions, in your agent prompts, and use the recorded reads to audit what an answer was based on.
Vulnerability reports
Found something? Talk to us — reports go straight to the engineers who operate the services.
Changes to this page
This page is versioned in git and dated; material changes update the date above.