Repository navigation
Copilot DLP and Governance
Copilot → Copilot DLP and Governance
Microsoft Purview DLP policies can stop Copilot using content. The DLP impact report (Insights → DLP impact) shows which agents and which people were affected, how often, and by which policies — agent and person rows expand to reveal the individual policies behind their numbers.
This needs no extra permission and no extra import toggle. It works with the
Copilotimport you already have on.
That is not an accident of design, it is how Microsoft publishes the data. DLP policies scoped to the "Microsoft 365 Copilot and Copilot Chat" location do not emit records on the separate DLP.All audit feed — Microsoft documents that feed's workloads as Exchange Online, Endpoint and SharePoint/OneDrive only. A Copilot block is reported inside the CopilotInteraction record instead, on AccessedResources[].Status and AccessedResources[].PolicyDetails, alongside AgentId / AgentName. Those records are therefore the only source that can attribute a DLP block to a specific agent. See Audit logs for Copilot and AI applications.
Every figure is split into two columns, and the difference matters:
| Column | Meaning |
|---|---|
| Blocked | The user did not get the content. |
| Audited only | A policy matched but withheld nothing — typically a rule still in simulation (RuleMode = "Audit only"), or one whose actions are notify/report only. |
A rule can list BlockAccess and still block nothing because it is running in simulation, so the two are never merged. The classification is deliberately conservative in both directions: an access failure carrying no policy detail is treated as not DLP at all (it is usually a deleted file or a permissions problem), and an action we don't recognise is treated as non-blocking rather than inventing a block.
Since Stable build 1860 (PR #866, PR #891): for readers without See PII, agents, policies, sensitivity labels and tenant DLP policies that affected fewer than 5 people aren't listed, and neither is a policy under an agent that affected fewer. When a table, or an agent's list of policies, leaves one out, or a table has more rows than the 20 it lists, the row of the fewest people still listed is left out too, so the totals can't give the hidden one back: a table with more than 20 rows shows these readers at most 19. A table left with nothing to list says that rows were left out, instead of Nothing in this period. See Portal permissions → Small groups withheld from readers without See PII.
Available in 1852. These are audit-derived safety/grounding observations, not a prompt-text classifier or a compliance guarantee.
Below the DLP tables, the page shows the prompt-safety and grounding signals Microsoft records on every Copilot interaction, for the same period and narrowed by any filter an administrator has applied to your reports, like the rest of the page. After applying the normal release migration chain, it needs no additional permission, toggle or migration: every value comes from the Copilot audit records the Copilot import already stores. Everything shown is an aggregate, so it needs no portal role beyond signing in.
Here interaction means a counted turn, not each raw audit row. The counted record anchors its date and user scope. Signals, models and plugins from all its duplicate constituents are retained, including evidence just outside a date boundary; an inside echo does not pull an outside counted turn into the period. Identical resource accesses are reduced on their full identity (resource id, name, site URL, type, sensitivity label, action and list-item id), with their XPIA flags combined before counting. A positive flag on any constituent survives, explicit false is reported clean, and all-missing flags stay unknown.
| Figure | What it means | Stored in |
|---|---|---|
| Jailbreak attempts | Interactions with a prompt Microsoft flagged as an attempt to get round Copilot's safeguards, per 10,000 of the interactions on which the flag was reported. |
copilot_event_messages.jailbreak_detected (Messages[].JailbreakDetected) |
| Cross-prompt injection (XPIA) | Interactions in which Microsoft detected instructions planted in a resource Copilot used, per 10,000 of the interactions on which the flag was reported. |
copilot_event_accessed_resources.xpia_detected (AccessedResources[].XPIADetected) |
| Labelled content used | The share of the resources Copilot used that carried a sensitivity label. A resource counts once for each interaction that used it, so this measures how often Copilot works with labelled content, not how much labelled content exists. | copilot_event_accessed_resources.sensitivity_label_id |
| AI models | Interactions per model name, as a share of all interactions in the period. DEEP_LEO is the model Microsoft reports when Copilot used deep reasoning. |
copilot_ai_models / copilot_event_ai_models
|
| System plugins | Interactions per plugin, as a share of all interactions. BingWebSearch means Copilot searched the public web. |
copilot_ai_system_plugins / copilot_event_ai_system_plugins
|
| Agents with DLP blocks | The distinct agents a DLP policy blocked in the period, and the five blocked most. Whether each one is still worth keeping is its Retire / Review verdict on Copilot Adoption → Agents, which the card links to rather than repeating. |
copilot_dlp_events (the same data as the tables above) |
Model and plugin names are shown exactly as Microsoft reports them, in every portal language.
Each rate is shown with its numerator and its denominator, plus how many of the period's interactions reported the flag at all — for example "3 of the 12,000 interactions where Microsoft reported the signal" and "Reported on 12,000 of 20,000 interactions in this period (60%)". An interaction whose flag is missing is counted in neither figure: a missing flag means Microsoft did not report it, not that nothing happened, and counting it as clean would dilute the rate with interactions nobody checked. When no interaction in the period reported a flag, the card says Not reported instead of showing a rate, never 0 per 10,000.
Expect the reported count to be well below the number of interactions when the period reaches back before your upgrade to Stable build 1832, the first build that stored the two flags. Older interactions keep NULL (there is no backfill; the values do survive in audit_events.event_data if you ever need them), so they fall out of both sides of the rate rather than pulling it towards zero.
The model shares do not add up to 100%: Microsoft does not name the model for most Microsoft 365 Copilot interactions (Audit logs for Copilot and AI applications, ModelTransparencyDetails), and one interaction can use several models or plugins.
No index carries the jailbreak or XPIA flag. Loading this section reads copilot_event_messages and copilot_event_accessed_resources and joins them to the period's interactions. The turn-aware queries in 1852 also retain the evidence from each constituent record while counting a paired turn once.
Historical benchmark, not a timing guarantee for the 1852 queries: before turn pairing, a synthetic tenant of 200,000 users (4 million interactions over 180 days, 8.4 million messages and 7 million accessed resources) took about 6 seconds for a 28-day period and 11 seconds for 90 days on single-core SQL Server Express LocalDB. Those measurements do not establish the performance of the new canonical-turn queries or of your deployment. Query plans, hardware and retained history affect the result; measure the released version on a synthetic or approved test copy before setting timeouts or capacity expectations. Retaining more Copilot history can increase the cost even when the selected period stays the same.
To keep that off the rest of the page:
- The section is loaded by its own request: the DLP figures appear without waiting for it, and if it fails it shows its own error while the rest of the page carries on.
- Its result is cached in the web app for 10 minutes per period and per filter, so a reload, or a second administrator opening the page, costs nothing.
For the same reason these figures are not in the Copilot Adoption workbook's Snapshot facts sheet: that analysis runs on every Copilot Adoption load and has to stay quick.
When Purview cannot finish a DLP evaluation in time it marks the interaction as deferred (DLPEvaluationDeferred). That interaction's DLP outcome is unknown — neither blocked nor allowed — so it is not counted in either column. The portal does not show deferred evaluations and nothing about them is stored. The Office 365 importer counts them as it imports and logs the total with each batch it commits, for example:
DLP commit: 12 Copilot policy match(es) of which 4 blocked, 0 DLP.All rule match(es), 3 interaction(s) with a deferred (unknown) DLP evaluation.
A non-zero count that keeps appearing means the DLP figures for those periods are an undercount. Search the importer's WebJob log, or its Application Insights traces, for deferred (unknown) DLP evaluation.
The page's second section shows tenant-wide DLP activity (Exchange, SharePoint/OneDrive, endpoint devices) from the DLP.All feed. That part is optional and needs ImportDlp plus the separate ActivityFeed.ReadDlp consent — see Prerequisites → Data Loss Prevention import is optional.
It is shown in its own section, and is deliberately never combined with the Copilot figures: those records carry the literal UserKey = "DlpAgent" and no agent identity at all, so joining them to Copilot activity would imply an attribution the audit data cannot support.
Since Stable 1856 (PR #768); operator-only and opt-in. Earlier imports could save a DLP.All detection under the workload where it happened, creating SharePoint or Exchange activity metadata and inflating file/mail activity. Correct routing fixes new imports; it does not rewrite retained history.
The repair recognises only an exact documented DLP operation (DlpRuleMatch, DlpRuleUndo, DlpInfo) or an event with retained dlp_rule_matches evidence. It does not assume that an operation starting with Dlp is a policy event, or that audit_events.event_data retains the original RecordType. Unrecognised events without positive evidence are left untouched.
Before removing incorrect derived rows, it archives their native fields and resolved lookup values, including file names, URLs and Exchange object IDs, in the existing AnalyticsState table's DlpHistoricalArchive partition. This preserves context that might exist nowhere else. The original audit event, policy matches, shared dimensions and general-event JSON remain in SQL; only wrong file/mail metadata is removed and the general workload becomes DLP. There is no SQL schema migration or new permission beyond the existing Storage Table access.
-
Take a database backup and choose a bounded half-open UTC window. Run from the matching importer output folder with the deployment's normal configuration. Stop all importer instances while repairing; each batch briefly takes exclusive locks on the three workload-metadata tables, blocking activity metadata saves.
-
Create an access-controlled checkpoint folder outside every Git worktree, including junction/symlink aliases. The following dates and path are synthetic:
.\SPOInsights.WebJob.Office365ActivityImport.exe --repair-dlp-history ` '2000-01-01T00:00:00Z' '2000-02-01T00:00:00Z' ` 'C:\Private\Contoso\dlp-history.json'
-
Repeat the same command until
Scanned=0. Each invocation scans at most 500 positively identified events and saves its checkpoint only after successful correction. Archive writes run before SQL locks, with at most eight concurrent storage operations; execution time depends on SQL and Table latency and is not benchmarked. There is no whole-window lock or automatic production run. -
Restart the importer and compare the affected reporting period. A changed metadata snapshot aborts the batch with
DLP historical metadata changed after archival; retry after stopping the writer. Custom extra columns, incoming foreign keys/cascades or enabled triggers make the repair refuse rather than discard custom data. Archive verification failure or oversized metadata also leaves SQL unchanged.
Replaying a batch is safe. Changed context receives a separate content-addressed archive version rather than overwriting the original. If historical writers were active, run the window again using a new checkpoint: a row inserted behind a saved cursor is not found by continuing forward. A checkpoint is bound to its original target/configuration and window; do not reuse one for another deployment.
The archive is not an automatic rollback script and has no expiry. Treat it as tenant evidence with SQL-equivalent access controls. SQL retention and user-scope purge do not clear it; include the partition in your retention/erasure and backup procedures. Keep the backup if a full rollback is needed, and never attach the checkpoint or archive values to a public issue. See Runtime State Reference.
| Table | Purpose |
|---|---|
copilot_dlp_events |
One row per (blocked or audited resource × policy × rule) on a Copilot interaction. Links to copilot_chats, so the agent and user come with it. |
dlp_rule_matches |
Tenant-wide rule matches from the DLP.All feed. Links to audit_events. No agent identity. |
dlp_policies / dlp_rules / dlp_actions
|
Shared policy, rule (with severity and rule mode) and action lookups, used by both sources. |
Privacy: the solution records that a policy fired — policy, rule, severity, action and the affected resource's name and sensitivity label — and never the content that triggered it. It deliberately does not request the separate "Read DLP policy events including sensitive details" permission.
- Copilot analytics overview
- Copilot Adoption user guide
- Copilot Adoption setup (admins)
- Work Trend Index
- Copilot investment
- Copilot spend
- Agent costs
- Prompt categories
- Copilot credit history (preview)
- Agent 365 package catalog
- The Copilot data in depth
- Prerequisites
- App registrations setup
- Install with the installer
- Verify the deployment
- Security & permissions model
- Private endpoints (optional)
- Certificate authentication (optional)
- Manual installation
- Activity & usage analytics (Power BI)
- SharePoint page tracking