Security and Policy / Security and Policy
Security and Policy
Detect and control risk in agent sessions: risk analytics, policies, findings, shadow MCP inventory, and detection rules.
The Security and Policy section of the dashboard detects and controls risk in agent sessions. Policies scan session interactions for secrets, sensitive data, prompt injection, destructive tool use, and unsanctioned MCP servers, then flag or block what they find. Findings feed risk analytics, per-session review, and an inventory of shadow MCP servers discovered in agent traffic. Open it from the Security and Policy group in the project sidebar.
Access requirements
Section titled “Access requirements”Security and Policy pages require the org:admin scope, which the default Admin role holds and the Member role does not.
Sections
Section titled “Sections”The risk dashboard: findings clustered into ranked signals, with headline metrics, exposure by data type, a badge showing when risk analysis last ran, and suppression and exclusion actions. Watchdog replaces the former Risk Overview page.
The finding-level log: every policy finding across recently analyzed sessions, one row per finding, with redacted matches, filters, and a risk-focused session view for investigating any single finding.
Create and manage the policies that scan agent sessions. A stepped editor configures what to detect, which content is in scope, whether to flag, warn, deny, or quarantine, severity, and audience, including natural-language guardrail policies. Detection rules, exclusion rules, and session quarantines live here as tabs.
The inventory of AI tooling nobody provisioned centrally: the coding harnesses, assistants, and local models that enrolled devices report, with gateway access decisions for the tools that speak MCP, and the shadow MCP server inventory with its evidence-backed access reviews.
The catalog of built-in detectors behind guardrail policies (secrets, financial data, PII, healthcare, prompt injection, and more) plus custom rule authoring with CEL expressions, live testing, and AI-suggested rules. Lives as a tab on the Guardrails page.
The rules that suppress false-positive findings: expression criteria, global or per-policy scope, a built-in library of known-safe values, and one-click creation from a Watchdog signal. Lives as a tab on the Guardrails page.
Related organization settings
Section titled “Related organization settings”Additional security controls live under Organization settings rather than in the project’s Security and Policy group:
- Audit logs record every evidence reveal, policy change, exclusion edit, and access decision made on these pages.
- Roles and permissions (RBAC) define who holds the
org:adminscope these pages require, and supply the roles a policy audience or an access approval can name. - IDP and SSO connect the organization’s identity provider for sign-in and directory sync, which is where the department, group, and role attributes behind team and audience filters come from.
- API keys mint the hooks keys that instrumented agents use to send the session traffic these pages scan.
Authorization servers and live MCP session management live in the Identity section of the project sidebar, and per-person killswitches are managed on each identity’s Access tab under Identities.
Identity-level controls moved to the Identity group in the project sidebar: authorization servers for MCP authentication, MCP session management for active connections, and the Identities roster, where a person’s risk findings, devices, and killswitches are gathered in one place.