> ## Documentation Index
> Fetch the complete documentation index at: https://docs.blindsight.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Syslog export to your SIEM

> Send Endpoint policy events to your SIEM's syslog collector in CEF, JSON or RFC 5424 (key=value or structured data), over TLS, TCP or UDP.

Blindsight can send Endpoint events to your SIEM's syslog collector, one message
per event, as it happens. That covers every prompt the agent blocked, redacted or
flagged, every policy change, and every time a device loses or regains protection.

If you would rather pull events, use the API:
`GET /api/runtime-security/events` with an [API key](https://docs.blindsight.io/administration/api-keys.md)
that has the **Read endpoint events** permission, filtered by verdict and date.

## What is sent

Each message says what happened and where: the time, the workspace, the device,
the person, the AI service, the verdict, the detection categories, the reason code
and the severity.

How the device and the person are named follows your workspace's
[identity setting](https://docs.blindsight.io/endpoint/who-is-named.md):

| Identity setting | Device and person in the message |
| - | - |
| Named | The device's host name and the person's email address. |
| Pseudonymized | The same codes the console shows, for example `Person 7KQ2MD`. |
| Counts only | Single prompts are not sent. Policy and protection changes are sent hourly, without naming a device. |

The setting in force when a message is sent decides how it names people. This
includes a message that was waiting to be retried when the setting changed.

### What was typed is off by default

By default, no prompt, excerpt, file content or topic wording leaves Blindsight in
a syslog message.

To send the text, turn on **Include what was typed** for a destination. This needs
the **Send prompt text to a SIEM** permission, which administrators have, and a
destination that uses TLS. Each blocked, redacted or flagged prompt then carries its
text, exactly as the console shows it.

* **Only what your policy redacts is removed.** A redacted prompt arrives with those
  values replaced. On a blocked or flagged prompt, values your policy does not
  redact, such as names, companies and addresses, can still be readable. The
  original text, as the device sent it before redaction, is never sent.
* **Pseudonymized workspaces stay pseudonymized.** Email addresses, device names,
  account names and the names of people in the workspace are replaced in the text
  by the same codes the console shows.
* **Counts only sends no text.** Single prompts are not sent at all. If the setting
  was on, it stays on but paused, and resumes when the workspace names people again.
* **The text is read when the message is sent.** A prompt deleted by retention or an
  erasure request before then is sent without text.
* **A long prompt is shortened** to fit a collector's line, and marked as shortened.

Turning the setting on or off is recorded in the [audit trail](https://docs.blindsight.io/compliance/audit-trail.md)
with who did it. Pointing a destination that includes text at another collector
needs the same permission. An API key cannot turn it on. Turning it off needs only
**Manage alerts**.

> [!WARNING]
>
> Your SIEM keeps its own copy of the text, under its own retention and access
> controls, not this workspace's. Turn on **Include what was typed** only when your
> SIEM should hold it.

## Before you start

**1. Have a collector on a public address**

You need a syslog collector your SIEM reads from, reachable from the internet.
TLS on port `6514` is recommended.

**2. Allow the Blindsight address**

Allow this address in the collector's firewall:

| Region | Address |
| - | - |
| Blindsight cloud (EU) | `34.246.156.194` |

If the address changes, you are told in advance. A self-hosted installation
connects from its own network.

**3. Have your CA certificate ready**

If the collector's TLS certificate is signed by a private certificate authority,
have that CA certificate ready in PEM form.

Blindsight connects only to public addresses. A host name that resolves to a
private, loopback, link-local or cloud metadata address is refused when you save,
and again each time a message is sent.

## Set up a destination

You need the **Manage alerts** permission, which administrators have.

**1. Open Alerts**

Go to **Endpoint → Configurations → Alerts** and select **Send to a SIEM**.

**2. Where is your collector?**

Enter a name, the collector's host name and port, and the transport. Paste your
CA certificate if it is a private one. Select **Next**.

**3. What should it receive?**

Choose the format and tick what to send:

* Blocked, redacted and flagged prompts
* Policy changes
* Device protection lost, and device protection resumed

Tick **Include what was typed** only if your SIEM should hold the text of each
prompt. It is off by default.

**4. Test it, then start**

Select **Send test event**. The result appears on the spot: delivered, or the
reason it was not, such as a refused connection, no answer, an untrusted
certificate or no TLS on that port. Then select **Start sending**.

The test event is titled **Test alert** and names nobody, so it is safe to send to
a production SIEM. With **Include what was typed** on, it carries a line of example
text, so you can see which field the text lands in.

What you ticked becomes ordinary alert rules, listed under **Rules**, where you can
switch one off or change it.

### Delivery health

The destination's row shows its health: when it last delivered, the last error, and
how many events were dropped, if any.

Messages are sent in the background, never on the path that inspects prompts. A
slow or unreachable collector does not slow anyone down. A failed send is retried
several times over the following hours. The first message given up during an
outage is written to the audit trail. During a long outage, a destination drops new
messages instead of building an endless backlog, and counts what it dropped on the
destination's row.

Delivery is at least once, so a message can arrive twice after a retry. Every
message carries a delivery ID that you can use to remove duplicates.

## Transports

| Transport | Encrypted | Delivery confirmed | Use when |
| - | - | - | - |
| **TLS** | Yes | Yes | Always, unless your collector cannot do TLS. |
| **TCP** | No | Yes | Inside a private link you trust. |
| **UDP** | No | No | Only when nothing else is possible. A test over UDP can only confirm the message left. |

TLS is version 1.2 or later. The collector's certificate and its host name are
always verified, and verification cannot be turned off. Messages are framed one per
line (RFC 6587, non-transparent framing), which every common collector accepts on
TCP and TLS.

## Formats

All four formats carry the same facts. Choose the one your SIEM parses best. Every
message starts with an RFC 5424 header:

```text theme={null}
<PRI>1 TIMESTAMP blindsight rs-alerts - MSGID ...
```

* `PRI` is the facility (16, `local0`, by default) times 8, plus the syslog
  severity: 6 for low, 4 for medium, 3 for high and 2 for critical.
* `TIMESTAMP` is when the message was sent, in UTC. The event's own time is in the
  fields.
* `MSGID` is the event family for both RFC 5424 forms and JSON, and `-` for CEF.

### Event families

| Family | When |
| - | - |
| `endpoint_event` | The agent blocked, redacted or flagged a prompt, a response or a tool call. |
| `policy_changed` | A policy, group, detection or install setting changed. |
| `protection_lost` | A device stopped inspecting AI traffic. |
| `protection_resumed` | A device inspects AI traffic again after a pause. |
| `test` | A test event you sent. |

> [tab] CEF
>
> Recommended for Splunk, Microsoft Sentinel, IBM QRadar and ArcSight. The CEF
> record is the syslog message.
>
> ```text theme={null}
> <11>1 2026-11-06T10:15:41Z blindsight rs-alerts - - - CEF:0|Blindsight|Runtime Security|1|endpoint_block|Prompt blocked|8|rt=1793960130000 cat=endpoint_event act=block msg=A prompt to chat.openai.com was blocked. suser=dana@example.com dvchost=DANA-LAPTOP dhost=chat.openai.com reason=topic_rule deviceDirection=0 flexString1Label=categories flexString1=credit_card,topic_rule cs2Label=workspace cs2=Example cs4Label=deliveryId cs4=e1a7...
> ```
>
> | Header position | Value |
> | - | - |
> | Device Vendor | `Blindsight` |
> | Device Product | `Runtime Security` |
> | Device Version | `1` |
> | Signature ID | `endpoint_block`, `endpoint_redact`, `endpoint_flag`, or the family for other events. |
> | Name | For example `Prompt blocked`, `Response redacted` or `Policy changed`. |
> | Severity | 3 low, 5 medium, 8 high, 10 critical. |
>
> | Extension key | Meaning |
> | - | - |
> | `rt` | When the event happened, in milliseconds since 1970 (UTC). |
> | `cat` | The event family. |
> | `act` | For Endpoint events, the verdict: `block`, `redact` or `flag`. |
> | `msg` | A one-sentence summary. |
> | `suser` | The person, named as your identity setting allows. |
> | `dvchost` | The device, named as your identity setting allows. |
> | `deviceExternalId` | The device's ID in Blindsight. |
> | `dhost` | The AI service's host. |
> | `reason` | For a block, the reason code, for example `topic_rule`. |
> | `deviceDirection` | `0` for a prompt or tool call going out, `1` for a response coming back. |
> | `flexString1` (`categories`) | Detection categories, comma separated. |
> | `flexString2` (`provider`) | The AI provider. |
> | `cs1` (`rule`) | The Blindsight alert rule that sent it. |
> | `cs2` (`workspace`) | The workspace. |
> | `cs3` (`sourceId`) | The event's ID in Blindsight. |
> | `cs4` (`deliveryId`) | This delivery's ID, for removing duplicates. |
> | `cs5` (`text`) | What was typed, only when **Include what was typed** is on. |
> | `cs6` (`topicRuleId`) | The ID of the [topic rule](https://docs.blindsight.io/endpoint/topic-rules.md) that matched, when one did. |
> | `cn2` (`redactedOnDevice`) | How many values the device itself replaced. |
> | `cn3` (`textTruncated`) | `1` when the text was shortened to fit. |
> | `request` | A link to the event in the console. |
>
> Values are escaped as the CEF specification requires (`\=`, `\\`, and line
> breaks as `\n`), so any text in a value reads back exactly.

> [tab] RFC 5424 (key=value)
>
> A standard RFC 5424 message with no structured data element. Every field is a
> `key="value"` pair in the message, followed by the summary as `msg`. Splunk,
> QRadar, Sentinel and the syslog-ng `kv-parser()` split this form without
> configuration.
>
> ```text theme={null}
> <11>1 2026-11-06T10:15:41Z blindsight rs-alerts - endpoint_event - family="endpoint_event" severity="high" occurredAt="2026-11-06T10:15:30Z" workspace="Example" verdict="block" direction="input" person="dana@example.com" device="DANA-LAPTOP" identityMode="named" serviceHost="chat.openai.com" provider="openai" categories="credit_card,topic_rule" reason="topic_rule" eventType="endpoint.block" deliveryId="..." msg="A prompt to chat.openai.com was blocked."
> ```
>
> Values are in double quotes, with `\"` and `\\` escaped. A field with no value
> is left out.
>
> > [!TIP]
> >
> > If you include what was typed, prefer CEF or JSON. A prompt often contains
> > double quotes, and some SIEMs stop a value at the first escaped quote.

> [tab] RFC 5424 (structured data)
>
> The same fields in one RFC 5424 structured data element, with the summary as
> the message. The SD-ID is `blindsight@67580`, under Blindsight's IANA private
> enterprise number 67580, and it does not change. Choose this form when your
> collector reads RFC 5424 structured data, for example the syslog-ng
> `.SDATA.blindsight@67580.*` values or the rsyslog `structured-data` property.
>
> ```text theme={null}
> <11>1 2026-11-06T10:15:41Z blindsight rs-alerts - endpoint_event [blindsight@67580 family="endpoint_event" severity="high" occurredAt="2026-11-06T10:15:30Z" workspace="Example" verdict="block" direction="input" person="dana@example.com" device="DANA-LAPTOP" identityMode="named" serviceHost="chat.openai.com" provider="openai" categories="credit_card,topic_rule" reason="topic_rule" eventType="endpoint.block" deliveryId="..."] A prompt to chat.openai.com was blocked.
> ```
>
> Parameter names are the field names in the reference below. Values are escaped
> as RFC 5424 requires (`\"`, `\\` and `\]`), and a line break becomes a space.
> A field with no value is left out.
>
> The key=value form does not change: a destination that uses it keeps receiving
> the same shape. To move to structured data, edit the destination and change its
> format.

> [tab] JSON
>
> The same fields as one JSON object in the message, plus `title`, `summary` and
> `consoleUrl`.
>
> ```text theme={null}
> <11>1 2026-11-06T10:15:41Z blindsight rs-alerts - endpoint_event - {"family":"endpoint_event","severity":"high","occurredAt":"2026-11-06T10:15:30Z","workspace":"Example","verdict":"block","direction":"input","person":"dana@example.com","device":"DANA-LAPTOP","serviceHost":"chat.openai.com","provider":"openai","categories":"credit_card,topic_rule","reason":"topic_rule","eventId":"...","deliveryId":"...","title":"Prompt blocked","summary":"A prompt to chat.openai.com was blocked."}
> ```

### Field reference for RFC 5424 and JSON

The same names are used in the key=value form, as structured data parameters and as JSON keys.

| Field | Meaning |
| - | - |
| `family` | The event family. |
| `severity` | `low`, `medium`, `high` or `critical`. Blocks are high. Redactions and flags are medium. |
| `occurredAt` | When the event happened, in UTC (ISO 8601). |
| `workspace`, `workspaceId` | The workspace. |
| `verdict` | `block`, `redact` or `flag`, for Endpoint events. |
| `direction` | `input` (prompt), `output` (response) or `tool_call`. |
| `person` | The person who sent the prompt, as your identity setting allows. |
| `device`, `deviceId` | The device, as your identity setting allows. |
| `identityMode` | `named` or `pseudonymised`. |
| `serviceHost`, `provider`, `model` | The AI service the prompt went to. |
| `categories` | Detection categories, comma separated. |
| `reason` | For a block, the reason code. |
| `topicRuleId` | The ID of the [topic rule](https://docs.blindsight.io/endpoint/topic-rules.md) that matched, when one did. |
| `redactedOnDevice` | How many values the device itself replaced. |
| `rule` | The Blindsight alert rule that sent it. |
| `eventId`, `eventType` | The event in Blindsight. |
| `deliveryId` | This delivery's ID, for removing duplicates. |
| `text` | What was typed, only when **Include what was typed** is on. |
| `textTruncated` | `true` when the text was shortened to fit. |

## SIEM examples

> [accordion] Splunk
>
> With Splunk Connect for Syslog (SC4S), listen with TLS on `6514` and choose
> **CEF** in Blindsight. SC4S recognizes the records by their vendor and product.
>
> ```text theme={null}
> index=* sourcetype=cef "Blindsight"
> | search act=block
> | stats count by suser, dhost
> ```
>
> Without SC4S, a TCP or TLS data input with source type `syslog` also works. With
> **JSON** chosen, use `| spath input=message`, or extract the JSON after the
> header, to read the fields.

> [accordion] Microsoft Sentinel
>
> Use the **Common Event Format (CEF) via AMA** data connector with a Linux log
> forwarder, and choose **CEF** in Blindsight. Events land in the
> `CommonSecurityLog` table.
>
> ```kusto theme={null}
> CommonSecurityLog
> | where DeviceVendor == "Blindsight" and DeviceProduct == "Runtime Security"
> | where DeviceAction in ("block", "redact", "flag")
> | project TimeGenerated, DeviceAction, SourceUserName, DeviceCustomString2, DestinationHostName, FlexString1
> ```
>
> `DeviceCustomString2` is the workspace and `FlexString1` holds the detection
> categories.

> [accordion] IBM QRadar
>
> Add a log source with these settings:
>
> * **Log source type**: Universal CEF
> * **Protocol**: TLS Syslog on port `6514`, or Syslog
> * **Log source identifier**: `blindsight`
>
> Choose **CEF** in Blindsight. Map `endpoint_block`, `endpoint_redact` and
> `endpoint_flag` to QRadar events in the Universal CEF event mapping, so you can
> use them in rules and offenses.

## Troubleshooting

> [accordion] Nothing arrives
>
> Send a test event from the destination's dialog. It names the problem. A
> refused connection is usually the port or the transport. No answer is usually a
> firewall that does not allow `34.246.156.194`. An untrusted certificate needs the
> CA certificate added to the destination.

> [accordion] The test arrives but no events do
>
> Check under **Rules** that the rules for this destination are on. Check that
> your workspace does not show counts only, because single prompts are not sent
> then.

> [accordion] You want the prompt text in the SIEM
>
> Edit the destination, tick **Include what was typed** on the second step, and
> save. Messages sent from then on carry the text.

> [accordion] You changed the collector
>
> Edit the destination and send a test event before you save.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.