Security
How Trailboss keeps workspaces apart, decides who may do what, and keeps a record you can trust. To report a vulnerability, write to [support@trailboss.dev](mailto:support@trailboss.dev) with "security" in the subject.
Workspaces are isolated
Every record belongs to one workspace, and the database enforces it: each request runs with its workspace set, and row-level security hides every other workspace's rows. Automated tests check that every table is covered.
Signing in
Email links, GitHub, authenticator apps and passkeys. Sessions are kept in secure, HTTP-only cookies. Risky approvals and production changes ask for a second factor again, within minutes, so a session alone is not enough.
Roles and approvals
Owners, admins, engineers and viewers each see and change only what their role allows. Approvals follow risk: the riskier the change, the more it needs, and the person who asked can't be the one who approves when your policy says so.
API tokens carry only the scopes you give them, expire, are shown once, and are stored as hashes; Trailboss keeps the last four characters so you can tell them apart.
The audit log
Every action, by a person, an agent or Trailboss itself, is recorded in an append-only log where each entry carries the hash of the one before. Batches of it are also archived to write-once storage, so changing history would be detected.
Agents work inside limits
Jobs are signed, and checked against your plan, budget and policies before they run. Agents act through policies and approvals, not open access: what they can touch is decided per job. Text agents read from repositories, issues and your knowledge sources is handled as untrusted reference material.
What a release is checked for
Before a release can be approved, a separate job checks out the code and runs: Gitleaks (secrets), Semgrep (static analysis), Trivy (known vulnerabilities in dependencies), Checkov when the repository has infrastructure code, and the repository's own tests. A check that was required and didn't run fails the release; it never counts as a pass.
To have your integration and end-to-end suites run there too, declare them in .trailboss/checks.json:
{
"integration": "npm run test:integration",
"e2e": "npx playwright test"
}Each runs in the same sandbox as your tests and becomes its own gate on the release. Without the file those two gates say "not declared" rather than passing.
How the running app behaves
Every preview is also probed while it runs, whatever framework or library it was written with: security headers, whether session cookies are HttpOnly and SameSite, error pages that show stack traces, and links that redirect to another site. From medium risk, a serious result fails the release's security gate and a lesser one is shown as a warning.
To have tenant isolation and cross-site request protection checked as well, declare two fixture users from different tenants in .trailboss/probes.json. Your seed step creates them; they never exist in production.
{
"users": [
{ "name": "a", "login": { "path": "/api/login", "json": { "email": "a@fixture.test", "password": "fixture" } }, "owns": ["/api/orders/1001"] },
{ "name": "b", "login": { "path": "/api/login", "json": { "email": "b@fixture.test", "password": "fixture" } }, "owns": ["/api/orders/2001"] }
],
"mutations": [{ "method": "POST", "path": "/api/orders", "json": { "item": "rope" } }]
}Each user signs in and reads what it owns; then each tries the other's records, an anonymous caller tries them too, and the state-changing request is sent the way another site's page would send it. One tenant reading another's record blocks the release. Apps that sign in with a token instead of a cookie add "token": "data.token" to the login, naming the field of the answer that holds it. Without the file, those two checks say "not checked" rather than passing.
The Standards page lists every requirement agents are given. Each one names the check that verifies it, or says "Guidance, not verified" when nothing automatic does. Trailboss doesn't claim a tool looked at your code when it didn't.
Your data
- Previews use generated sample data, never production data.
- Logs and job output have secrets removed before they're stored.
- Owners can export everything the workspace holds, and can delete the workspace: after a 30-day grace period, every record is removed.
The application
Strict security headers and a content security policy on every page, limits on how often anyone can call the API or try a credential, and checks that stop the server from being pointed at internal addresses. Dependencies, container images and the code history are scanned for known vulnerabilities and leaked secrets on every change.