Scope and boundaries

What ARIA governs

• Agents and workflows built and run on ARIA or Hermes (see Agents, workflows, and MCP)
• External agents and orchestrators that submit actions to ARIA for a decision, and report the outcome
• The evidence trail for both: decisions, approvals, audit records, and replay

What ARIA does not do

• It does not discover unmanaged AI. ARIA governs agents that are registered with it or that call it. It does not scan a network or SaaS estate for agents nobody told it about. Pair it with discovery tooling if shadow AI is a concern.
• It does not promise to stop prompt injection. ARIA labels content agents read as untrusted, scans it, and holds risky actions in a run that read something suspicious. Detection is imperfect, and the control is the hold. It does not cover the in-app ARIA Assistant chat, direct calls to model APIs that bypass ARIA agents, or images, and Hermes does not yet include these controls. See Prompt-injection defense.
• It is not a data loss prevention product. Policy rules can match classification, data residency, and simple personal-identifier patterns, sensitive fields are redacted from audit records, and a flagged run's outbound calls can be limited to an allow-list. Use dedicated DLP for content inspection across your environment. See Policy and approvals.
• It does not red-team your models. Agent test suites support regression testing, and ARIA does not run adversarial testing or vulnerability scanning of models.
• It does not certify a deployment. ARIA supports authorization. Your authorizing official grants it. See Identity and compliance alignment.
• It does not guarantee that a model is right. ARIA requires evidence, records decisions, and supports review. Model output still needs human judgment.

How ARIA fits with other platforms

Organizations rarely have one AI stack. ARIA can serve as an independent point of policy, approval, and evidence beside orchestration platforms, agent frameworks, security tooling, and your existing security operations. Integration details depend on the platform, so validate telemetry access, enforcement points, and responsibility boundaries in a technical discovery engagement.