Guardrails
An agent that can read can be told what to do by whatever it reads. A web page it fetches, a ticket it summarizes or a shared document can carry instructions aimed at the agent, not at you. An agent can also act: send mail, post to a channel, write to a system or run a command. Then content it merely read can become an action you never asked for. This is indirect prompt injection, and input guardrails exist to contain it.
HYPR AgentPass™ input guardrails do not try to out-argue a hostile instruction. They watch what a session has done: what it read, how sensitive that was, and where it now tries to send it. When that combination is dangerous, they hold or stop the call, however the content was worded.
In the console, Guardrails and each of its pages carry a Preview tag.
Input guardrails are enforced at the gateway, alongside your policies and tool access. They apply to every kind of connector a session touches, whether an MCP server, a REST API or an AI provider, so a session cannot sidestep them by switching tools.
The Pieces
Input guardrails are a set of controls that work together. Each has its own page.
| Control | What it does | Read more |
|---|---|---|
| The Rule-of-Two breaker | Holds or stops a call when a session that read untrusted and sensitive data reaches a way out. A held call raises an approval that carries its evidence | The Rule-of-Two Breaker |
| Tool-definition integrity | Flags a tool whose definition changes after it was approved, and raises an alert. With quarantine turned on, agents also lose the tool until a person approves the change | Tool-Definition Integrity |
| Alerts and containment | Turns a guardrail enforcement into a deduplicated, budgeted alert, gives security operations a triage queue, and lets a responder contain the agent behind an alert | Alerts and Containment |
| Posture and history | Shows what the guardrails have done and what still waits on a person, and trends that activity over time | Posture and History |
| The session investigator | Reads Session watch findings and, once you let it, pauses or terminates a coding-agent session | The Session Investigator |
| Bring-your-own judges | An optional layer that asks a model you choose about the content behind a held or stopped call | Bring-Your-Own Judges |
| The Proving Ground | Runs a scripted attack pack against an isolated sandbox, and shows which attacks the guardrails held or stopped, off versus on | The Proving Ground |
| Settings | Alert volume, containment, tool quarantine, approval timeouts and the session investigator | Guardrail Settings |
In the console, these pages sit under Guardrails in the sidebar: Posture, Alerts, History, Proving ground, Session investigator, Judges and Settings. You see the entries your permissions allow. See Who Manages Guardrails.
How the Breaker Reasons
Three facts about a session decide whether the breaker acts. You do not compute them. The gateway tracks them as the session runs.
- Trust. Every connector has a Source trust setting. It is Untrusted by default. Mark a connector Trusted, with a recorded reason, only when its content comes from a source you control. One read from an untrusted source marks the session untrusted for the rest of its life.
- Sensitivity. Tool reads carry a sensitivity tier from your data classification. A session's sensitivity only rises. Reading Confidential data once keeps the session at that level.
- A way out. Each tool has a sink: external communications, a state change, a credential or code execution. A read is not a sink; sending an email is.
When a session has read untrusted and sensitive data, and then reaches an external-communications or state-change sink, the breaker acts. The Rule-of-Two Breaker explains what happens next.
Turning Guardrails On
You adopt input guardrails deliberately, so you can see what they would do before they do it. Both steps need Data labels: Manage, which Admin holds out of the box.
- Turn on session tracking. The breaker needs to know what each session has read. Turn it on from Settings → Data classification. With it off, the
session.*facts the breaker relies on are absent, and the guardrail rules never match. See Session Tracking. - Adopt the guardrail starter policies. Open Settings → Data classification, and click the Policies tab. It lists four
guardrail-starters beside the classification and session starters. Preview matches shows how many calls a starter would have matched in the last seven days. Add this policy creates it in observe mode, so it records what it would have held or stopped without changing traffic. Switch it to enforce once you trust it. See How Policies Are Evaluated.
The Policies tab of Settings → Data classification now shows each guardrail- starter as Already added.
Alert, containment, approval, tool-integrity and session investigator behavior is set on Guardrails → Settings. See Guardrail Settings.
Who Manages Guardrails
What you see under Guardrails follows your permissions.
| Permission | Lets you | Out of the box |
|---|---|---|
| Guardrails: View | Open Posture, Alerts, History, Proving ground and Session investigator, and export evidence | Admin, Operator, Security analyst, Auditor and Model governance owner |
| Guardrails: Manage | Work alerts, contain agents, run the attack pack, mark a session pause a false positive, and open Settings | Admin and Security analyst |
| Settings: Edit, with Guardrails: Manage | Save guardrail settings and per-agent overrides | Admin |
| Model governance: View | Open Judges | Admin, Security analyst, Auditor and Model governance owner |
| Model governance: Manage | Register, promote and bind judges | Admin and Model governance owner |
Security analyst triages and contains, and reads broadly, but cannot change policies, settings, connectors or credentials. Auditor reads and exports, and changes nothing. Model governance owner governs judges. See Roles for Security Operations, The Model Governance Owner and Permissions by Role.
Related
- Policies and Guardrails — the governance controls that guardrails sit alongside
- Session Watch — a complementary, observational review of agent sessions
- Roles and Permissions — assigning the guardrail roles