Skip to content

Security

Open it: Security · under Security & audit in the console menu. For SuperAdmin and SecurityAdmin only; anyone else sees a note saying so.

Every call A3T makes on someone’s behalf (a tool, a model, a flow start, a deploy, a change to a resource) is checked at its choke point against Cedar policies, as the person it is for, with their role in the project the call touches. Anything not permitted is denied, and a forbid beats any permit. Security is where those policies are written, tested and read.

Items lists:

ItemWhat it is
IdentityThe sign-in provider and its users, platform-wide
A policy per project<project>/<name>: rules that apply to that project’s calls only
baselineThe deny-by-default rules compiled into A3T. Read only; they apply everywhere, beneath every project policy
  1. The call’s principal is the person (or the agent acting for one), with their role in the project the call touches: write access makes them Admin there, run access Operator, read access Viewer. SuperAdmin and SecurityAdmin keep their platform roles.
  2. The baseline and the project’s own policies are evaluated together. Platform policies (in a3t, system or default) apply to every project.
  3. Nothing permits it, or a forbid matches: denied. The refusal names every policy that refused it, and, for a baseline rule, what access it needs.
  4. The decision is written to the Decision log.

A refusal reads, for example:

Refused by policy qa-operator-support/no-flow-deletes: mcp:call on Tool 'delete_flow'.

The baseline: what it permits and forbids, compiled into the runtime

Each baseline rule has a name, baseline/<rule>, used in refusals and the log. The main ones:

RuleWhat it says
baseline/humans-run-stepsPeople may invoke tools, call MCP tools and models, and start flows
baseline/agents-run-for-a-userAgents may too, but only when acting for a person
baseline/destructive-toolsTools that delete or change data need the Admin role (or SuperAdmin) in the project, as a person, not an agent
baseline/data-writesWriting or deleting data through a database step needs Admin, as a person
baseline/flow-exposed-tools, baseline/flow-enabled-toolsA flow may only use the tools it declares
baseline/policies-are-infosecsOnly SecurityAdmin and SuperAdmin may write Cedar policies
baseline/authoring-needs-adminCreating flows, agents or templates through tools needs Admin in the project
baseline/prod-deploy-needs-admin, baseline/isolated-deploy-needs-adminProduction and isolated placement need Admin, as a person
baseline/prod-operate-needs-adminPausing or stopping production needs Admin

Policy shows its full Cedar text.

Pick a project in the top bar, then New policy.

New policy: name, project, what it is for, and its Cedar

FieldWhat it meansExample
NameWhat it doesno-flow-deletes
ProjectThe project whose calls it governs. A policy in a3t, system or default governs every projectqa-operator-support
What it is forOne sentence, shown on its OverviewNobody in this project deletes flows through tools
CedarThe rules: permit and forbid statementsbelow
// Nobody in this project may delete a flow through a tool call.
forbid(principal, action in [Action::"tool:invoke", Action::"mcp:call"], resource)
when { resource has name && resource.name == "delete_flow" };

Create and apply compiles it first: a policy that doesn’t compile is refused with Cedar’s error, and nothing changes. Once saved it applies at once, on every copy of FlowR.

A policy's Overview: what it permits and forbids, and where it applies

TabWhat it shows
OverviewWhat it does, counts of permit and forbid rules, the actions and roles it names, where it applies; Delete (its project falls back to the baseline)
PolicyIts Cedar text, to edit. Save compiles it before saving and applies it at once
TestA dry run (below)
DecisionsRecent decisions that name it

Test: role, action, resource; the decision and the policies that decided

FieldWhat it means
RoleWho asks: SuperAdmin, SecurityAdmin, Admin, Operator or Viewer
Actionmcp:call, tool:invoke, resource:create, resource:update, resource:delete, llm:call, flow:start, flow:deploy
Resource type and ResourceWhat it is done to: a Tool, Resource, Llm, Flow or Runtime, by name, e.g. delete_flow
With this policy as written hereTest the text in the editor before saving it

Decide shows Allowed or Denied, Why, and Decided by: every policy that decided it. Nothing runs and nothing is recorded. In the example, delete_flow is denied by qa-operator-support/no-flow-deletes, and list_flows is allowed by baseline/humans-run-steps.

An agent’s own Audit › Policies has the same test, for that agent and its tools.

Identity: the sign-in provider, and how its roles become A3T roles

SectionWhat it shows
Sign-in providerIts state, realm and number of users; Manage opens the provider. Users lists them
RolesHow the provider’s realm roles become A3T roles at sign-in
A3T roleWhat it may do
SuperAdminEverything, in every organisation, including Security and the Decision log
SecurityAdminSecurity and the Decision log: policies, identity, every decision
AdminRuns projects; does not see Security or the Decision log
OperatorWorks in their projects
ViewerReads only, everywhere

An organisation can sign its people in with its own identity provider; see Settings.