Violations
Overview
Section titled “Overview”Govern → Policies & rules → Violations lists every time a Shield360 policy acted, whichever part of Shield360 made the decision:
| Category | What it is | Where it is decided |
|---|---|---|
| Sensitive data | Personal data, credentials or keys found in a prompt, file or command | Browser extension, AI proxy (self-run and hosted), coding-agent hooks, SDK |
| AI tools | An AI site or tool that is blocked or needs a warning | Browser extension, AI proxy |
| Unapproved models | A request for a model or provider outside the environment’s approved list | AI proxy (self-run and hosted) |
| Sequence policies | A risky order of agent actions, such as a secret read followed by a network call | Coding-agent hooks, MCP gateway, the scan of traces |
| Prompt injection | Text trying to take over an agent | Coding-agent hooks, MCP gateway |
The specialised pages stay: Detect & respond → Detections → Guard events for sensitive data and AI-tool decisions, and Detect & respond → Detections → Prompt injection for detections. Each links to the feed, and the feed links back to them.
Outcomes
Section titled “Outcomes”| Outcome | Meaning |
|---|---|
| Blocked | Never sent or run. |
| Masked | Sent with the sensitive data removed (reported by guards as redact). |
| Warned | Went ahead, and was recorded. |
| Sent anyway | The person was warned, then sent anyway with a reason. The reason is shown with the event. |
| Block not enforceable | The policy was set to block, but the agent offers no way to refuse at that point, so it was recorded as a warning. This is the same rule as on the sequence-policy and prompt-injection pages. |
| Allowed | Checked and let through. Only reported where a site asks for it. It is listed, but not counted as a violation. |
The cards show Violations (everything but allowed), Blocked, Masked, Warned, Sent anyway, Not enforceable and People affected (distinct people across every category shown).
Filters
Section titled “Filters”- Category: sensitive data, AI tools, unapproved models, sequence policies or prompt injection.
- Outcome: any of the outcomes above.
- Sources: Browser extension, AI proxy, Coding agents, SDK, Gateway agents, or the scan of traces.
- Severity:
- Sequence policies use the policy’s own severity.
- Sensitive data takes the highest severity of its kinds of data. It uses the defaults Shield360 applies to sensitive-data findings: keys, credentials, card and national ID numbers are critical, and contact details are warnings.
- AI-tool decisions are warnings. Prompt-injection detections are critical.
- Policy: a sequence policy. Choosing one shows only that policy’s violations.
- Select a person or a kind of data in a row to narrow the feed to it.
The time range is the one in the page header. The newest 50 events are shown first; Load more continues from there.
Details
Section titled “Details”Select a row to open the same detail its own page shows:
- Sensitive data and AI tools open the Data protection event, with each value’s type, fingerprint and matching vault secret (never any character of the value), how often each value has turned up and by how many people on how many machines, the rest of the session and the person’s other events. Show every event with this value opens Data protection, filtered to that value.
- Sequence policies open the violation’s timeline of matched events.
- Prompt injection opens the detection: rules, score, where it was found and the masked snippets.
Values are masked where they were found. The full value never reaches Shield360, and the feed never shows more than the specialised pages do.
Permissions
Section titled “Permissions”The Violations tab needs policies.read. Each other category also needs the permission its own page needs:
| Category | Also needs |
|---|---|
| Sensitive data, AI tools, unapproved models | compliance.dataprotection.view |
| Prompt injection | compliance.injection.view |
A role without one of these sees the feed without those rows, and a note says which categories are hidden. Every read is checked on the server and scoped to the current project and environment. Opening the feed is recorded in the audit log as data access.