Risk Events

The Risk Events page is the finding-level log: every policy finding across recently analyzed sessions and MCP tool calls, one row per finding. Open it from Security and Policy > Risk Events in the dashboard.

Where Watchdog clusters findings into ranked signals for triage, Risk Events lists the raw findings themselves — useful for auditing individual matches, tracking down a specific rule or user, or handing a single finding to a teammate.

Access requirements

Viewing this page requires the org:admin scope. Access is included in the default Admin role but not the Member role.

Event list

Each finding shows its timestamp, category, severity, rule, session or tool, user, matched content, and the guardrail policy that raised it. The list loads continuously with a running count of findings shown.

Filters narrow the stream. Policy, Date range, and MCP server are always visible, and the rest live behind More filters:

• Policy: including inactive policies, which are labeled and produce a notice
• Date range
• MCP server: findings on tool calls through one server, gateway endpoint, or Platform MCP toolset, with the finding count beside each option
• Rule ID: matches rule IDs containing the entered text
• User
• Unique matches only
• Assistant: including sessions with no assistant

MCP findings

MCP-scoped policies record findings on tool calls rather than on session messages. An MCP finding carries an MCP badge and the name of the server, gateway endpoint, or toolset in place of a session name, with the surface, tool, and outcome beneath it. The outcome is one of:

• Logged: a Flag policy recorded the finding and the call proceeded
• Denied: a Block policy rejected the tool request before it reached the server
• Withheld: a Block policy kept the tool result from the client

To list the findings for one server, filter by MCP server, or select Open in Risk Events on the server’s Guardrails tab.

Redaction and reveal

Matched content is redacted by default. Individual matches can be revealed inline, and a Reveal all toggle in the header switches the whole view. Findings from natural-language guardrails open a match dialog with the evaluated content. Redacted matches are visible with the org:admin scope alone; revealing them follows the same audit-logged rules as Watchdog.

Revealing a match also requires the chat:read scope: on the finding’s session for a session finding, and on all sessions for an MCP finding, which has no session. The matched content of an MCP finding is stored encrypted for 90 days; findings recorded before evidence storage show Evidence not stored.

Suppressing findings

Select findings with the row checkboxes (or Select all loaded findings) to act on them together. Suppress Once dismisses the selection without changing any policy, and Create Rule opens an exclusion rule prefilled from the selection. These are the same actions as the Watchdog suppress menu, applied to individual findings rather than signals.

Investigating a finding

Clicking a finding opens the finding drawer. Move to the next or previous finding with j and k, and use Copy link to hand the finding to a teammate. Suppress in the drawer header offers the same Suppress Once and Create Rule actions as the list.

For a session finding, Open transcript opens the session in a risk-focused view, the same transcript panel as Agent Sessions, with flagged messages highlighted and unflagged messages dimmed.

For an MCP finding, the drawer shows the tool call itself:

• Call path traces the call from the client through the gateway to the MCP server, and states what the gateway did: stopped the request, withheld the result, or forwarded the call unchanged.
• The payload section shows the tool request arguments or the tool result that was scanned. Select Reveal matches to unredact the flagged values in place. Revealing the full payload requires the chat:read scope on all sessions, and every reveal is recorded in the audit log. Payloads are stored when Record Tool I/O is on for the organization, up to 64 KB each, for 90 days.
• Findings in this call lists every other finding on the same call.
• Details lists the user, MCP server, tool, surface, phase (request or response), policy, rule ID, and timestamp. The user column names the person behind the session; their identity page gathers every finding, device, and session for that one person.

Findings can also be delivered to an OTLP destination through a Risk findings export, with exclusions re-checked before every delivery; see Data export.