Guardrails
Limit what keys and members may spend, which models and clouds they reach, and what sensitive data may leave in a prompt.
How guardrails apply
A guardrail is a named set of rules in one workspace. Every workspace has a default guardrail that applies to every key in it; owners and admins can edit it, and it is never deleted. A named guardrail adds its rules on a key (one per key) or on a member for one workspace (one per member per workspace). Owners and admins create, edit, delete and assign guardrails on the console's Guardrails page; members can read them and the combined rules of their own keys.
A request must pass the workspace default, the key's guardrail and the member guardrail of the person who created the key. The stricter rule wins:
| Rule | How guardrails combine |
|---|---|
| Models and clouds | Only what every applicable guardrail allows. Lists with nothing in common allow nothing. |
| Spending limits | Each limit counts on its own; the first one reached stops the key. |
| Sensitive data | Block beats redact, and redact beats off, per item. |
| Prompt injection | Blocked if any guardrail blocks it. |
| Storage and export | "Never store or export prompts and replies" wins over the workspace settings for content. |
Spending limits
A guardrail can set a limit per day, per week and per month, in USD. The windows are UTC calendar periods: the day resets at 00:00 UTC, the week at 00:00 UTC on Monday and the month at 00:00 UTC on the 1st. The workspace default and a key guardrail count each key's spend separately. A member guardrail counts the total across that member's keys in the workspace, so a new key adds no allowance.
Models and clouds
A guardrail allows any model or a list of models, and any cloud or a list of the "Served by" clouds the console shows, such as Alibaba Cloud or Novita. A cloud in the list covers every route served under that name, including routes added later. A request is refused only when no allowed route remains for the model.
Sensitive data
Rules read the text of chat completions, Responses and Messages requests: the messages of every role, including the system prompt, and text input parts. They also read the text of these media requests:
- image:
promptandnegative_prompt - video:
promptandnegative_prompt - 3D:
prompt - audio generation:
promptandlyrics - music bed and product motion:
prompt - video highlights and clips: the focus prompt
- video characters:
nameanddescription - video assets:
name - speech (text to speech): the
input - embeddings: the
input
Replies, images, audio and video files are not scanned. Transcription takes no request text, so there is nothing to read. Rerank, the lang/* capability routes and realtime sessions are not read, nor is the image footnote watermark text or a character's tags. For each item a guardrail is off, redacts or blocks. Patterns are fixed and maintained by BeLive; you cannot add your own.
| Item | What is detected | Redaction label |
|---|---|---|
| Email addresses | An email address | [EMAIL] |
| Phone numbers | Singapore 8-digit numbers with an optional +65, and international numbers in E.164 form | [PHONE] |
| NRIC / FIN | Singapore NRIC and FIN numbers with a valid checksum | [NRIC] |
| Card numbers | 13 to 19 digits that pass the Luhn check | [CARD] |
| IP addresses | IPv4 and IPv6 addresses | [IP] |
| API keys and secrets | Common API key prefixes, private key blocks and similar credentials | [SECRET] |
Redact replaces each match with its label before the request is sent to a provider, stored or exported. The call succeeds, no response header marks it, and Activity tags the request "redacted" with the item types. Block refuses the request. For speech input and embeddings input, an item set to Redact refuses the request with 403 instead, because redacting would change what is spoken or embedded. Block refuses these requests too.
Prompt injection
Prompt injection is off or block. With block, a request whose text matches a short, fixed list of injection phrases, such as an instruction to ignore previous instructions or to reveal the system prompt, is refused. The check is a pattern match on the request text and adds no model call. It reads the same requests and the same text as the sensitive data rules: chat completions, Responses and Messages, and the media requests listed above.
When a guardrail blocks a request
A blocked request returns HTTP 403 with type guardrail_blocked, whether it is a chat or a media request. It is never sent to a provider and is not charged. The message names the guardrail and the reason, never the matched text:
{ "error": { "code": 403, "type": "guardrail_blocked", "message": "Blocked by the guardrail \"Client A keys\": Sensitive data: card. The request was not sent to a provider and was not charged." } }
The reason is one of a sensitive data item, a prompt injection pattern, a model or a cloud the guardrail does not allow, or a spending limit reached, for example Spending limit reached: $10 a day. A guardrail spending limit answers 403; only workspace and member budgets answer 402. Activity shows the request as "403 Blocked" with the guardrail and the reason. Do not retry a blocked request unchanged.
What is stored
When a workspace switches on Store in BeLive AI Hub on the Storage and export page, BeLive keeps the prompt text and request settings of each completed model request, and any text the model returns. A request that did not complete is not stored.
- Stored: the text you sent, such as a chat message, an image or video prompt, the text of a speech request or the input of an embeddings request, and the request settings, such as duration, resolution, aspect ratio or voice. For a transcription request it also keeps the transcript.
- Never stored: input files (images, audio and video) and generated images, audio, video and embedding vectors. Settings that count input files keep only the count.
- Redaction: the sensitive data rules of a guardrail read the text of chat and media requests, as listed under Sensitive data, and a request is stored after that redaction. Stored content and exports keep the redacted text and its labels, such as
[EMAIL]; the original is never stored. Text that no rule reads, such as the imagefootnote, is stored as it was sent to the provider. To keep requests out of storage, give their keys a guardrail that sets "Never store or export prompts and replies", or limit storage to other keys on the Storage and export page. - Kept: 90 days, encrypted, in Singapore, then deleted. Owners and admins read a stored request in the Activity drawer and can delete it. A guardrail that sets "Never store or export prompts and replies" stops storage for its keys.
For a video or audio generation job, while Send to your cloud is on, the prompt of a job that is still running is held until the job finishes, for at most 48 hours, so that your destination receives its record. This holds even if storage is switched off meanwhile. It is never added to stored content afterwards.
Send to your cloud delivers the same text as storage keeps, redacted or not by the same rule. A copy already delivered to a destination, a file you downloaded and an export written into your bucket are outside AI Hub: deleting stored content does not change them.
The Activity drawer says why a request is not stored. A request type that storage does not cover yet, and a request made before storage covered its type, show as not stored.
Export stored content
Owners and admins can take what a workspace stored as files, from Export stored content on the Storage and export page. It is there while Store in BeLive AI Hub is on. Choose a window, optionally some API keys and request kinds, and where the files go.
- Window: whole UTC days, the first and the last both included, at most 90 days. Stored content is kept for 90 days, so days before that are empty.
- Files: gzip JSON Lines (
.jsonl.gz), one file per UTC day and more than one for a day that is large. Each line is one request: the prompt text, request settings and any returned text, the same record Send to your cloud delivers. Input files and generated images, audio, video and embedding vectors are never exported. - Manifest:
manifest.jsonis written last. It lists the window and filters, every file with its record count, sizes and SHA-256, the totals, and how many requests were skipped because their content was deleted, was past 90 days, sat under a never-store guardrail or could not be read. It holds no content. - Download: BeLive keeps the files for 24 hours after the export finishes, up to 2 GiB compressed. Owners and admins of the workspace download them from the page while signed in, and each download is recorded in the audit log. An export that would be larger is refused; choose a shorter window or send it to a destination.
- Send to S3 destination: the files and the manifest are written into the bucket of an S3 destination that is not paused and has Include content on, while Send to your cloud is on. The bucket can be in any region, or on an S3-compatible service, and so outside Singapore. After that the content is in your cloud and BeLive does not reach into it.
- Limits: one export at a time in a workspace, and three exports a day.
A download export stays available for up to 24 hours after it finishes, even if one of its requests is deleted afterwards: a file made before the deletion still holds that request until it expires. Switching off storage or deleting all stored content removes the files BeLive holds at once, and an export that is running stops with the reason storage was switched off.
An export to an S3 destination keeps writing to the bucket it started with. Editing the destination while the export runs does not redirect it. To stop it, cancel the export or pause the destination.
For an S3 destination, BeLive keeps the access key you enter, encrypted, so that it can write to your bucket. Give it a write-only key scoped to one bucket prefix, not a key that can read or list the bucket. BeLive writes under ai-hub/, so for AWS an IAM user or role with s3:PutObject on arn:aws:s3:::your-bucket/ai-hub/* is enough to deliver. The connection test writes one small file, .connection-test, and deletes it again, so add s3:DeleteObject on arn:aws:s3:::your-bucket/ai-hub/*.connection-test only, or the test fails. Rotate the key with Replace on the destination.