Skip to content
Docs menu

Guardrails and filters in the console

Set the org's, key groups' and your own guardrails, test them on pasted text, and manage the prompt and response filters and the secret scanner from the Venom Cloud console.

View .md

Three console pages under Safety & policy decide what may leave your members' machines and what answers may carry:

Page What it is for Who can open it Who can change it
/guardrails the org's layered guardrails: the org's rules, each key group's and resold customer's, members' personal rule counts, and the resolved set for anyone audit:read (owner, admin, developer) org:manage (owner, admin) for the org; org:manage or that key group's admin for a key group
/me/guardrails every rule that applies to you, plus your own private rules any member of an org (org:read) you, for your own rules
/filters prompt and response filters (redact, block, warn, log) and the built-in secret scanner on answers audit:read org:manage

How the layers combine (strictest action wins, locks, Monitor mode, codes such as GR-ORG-PII-104) is described in Guardrails. This page covers the console screens.

Add an org guardrail

  1. Open /guardrails (operators first pick an organisation in the org switcher). The Org tab is selected.
  2. Choose Add guardrail, then a catalog entry (for example "Email addresses") or a blank form.
  3. Fill in the name, what it detects, whether it reads prompts, answers or both, the action, Enforce or Monitor first for N days, and the lock (locked, tighten, relaxable).
  4. Choose Test it: the draft runs together with the rules already in force on text you paste. Nothing pasted is stored.
  5. Save. The console answers "Saved. Every instance applies it within 10 seconds."

Each saved rule gets a code. When it fires, people see the layer, the rule's name and the code, never the matched text; the code's page is /g/<code> in the console.

The Guardrails page

/guardrails has four tabs, selected with ?tab=org, ?tab=teams, ?tab=members or ?tab=effective. The Teams tab is about key groups (a key group holds API keys, daemons, policies and GPU grants, not people; the API calls it team):

Tab Shows Controls
Org the org's own rules and the rules that live on the Filters, Budgets and Permissions pages (each links there). Per row: what it checks, action, scope, lock, and fires in 7 days (a rule in Monitor shows how often it would have fired) Add guardrail, edit, enable or disable, delete; the Drafting and prompt-injection classifier cards
Teams each key group's and resold customer's own rules, and the org's rules as that key group sees them a switch for a relaxable org rule, a stricter action or Monitor to Enforce for a tighten rule, nothing for a locked one; Clear returns to the org's setting. Pending member suggestions with Approve and Decline
Members' own who has personal rules, how many, how many org rules they turned off, how often theirs fired none: admins never see what a personal rule checks
Effective for… the resolved set for a key group, customer or member (?for=team:<id>, customer:<id>, member:<id>), in the order the checks run, with what was relaxed and which changes were refused Test text

Test text runs the effective set for the org, a key group, a customer or a member on pasted text. It returns the decision (block, keep_local or pass), each rule that fired with its layer and action, what monitored rules would have done, the rules that passed, and the text as it would be sent. Read it as: tool output checks the text the way the prompt-injection shield reads untrusted content.

Drafting a rule from a sentence

The add form can start from a sentence such as "Don't let anything with patient IDs leave my laptop". A model drafts the structured rule and fills in the form; you check each field and test it before saving. The draft is never saved on its own.

Setting (Drafting card, Org tab) Default Notes
enabled off an owner or admin switches it on; audited guardrail_drafting.updated
model the gear's classifier role, else this setting, else the prompt classifier's runs on the org's own provider keys, booked to the person drafting
daily_limit 20 drafts per person per day counted across the console and the CLI; past it, 429 guardrail_drafting_limit

A draft that fails the same checks as a hand-written rule (a regex the linear engine cannot run, an action the kind does not take) is refused with 422 guardrail_draft_unusable.

Prompt-injection classifier

The card of the same name sets whether a model double-checks text flagged by the prompt_injection detector, with enabled, model and daily_limit fields, saved by an owner or admin. The detector itself works without it; see Guardrails.

My guardrails

/me/guardrails lists every rule that checks your requests in one table:

  • the org's and your key group's rules, greyed, with where each came from. A relaxable rule has a switch you may turn off for yourself; a locked rule shows a lock; a tighten rule has neither;
  • your own rules, editable. They apply only to requests made with your own keys (a key your venom cloud login minted, or one you created as personal) and stay private: admins see that you have them, never their content;
  • Suggest to my key group on one of your rules sends a copy, with a note, to the key group's admin, who approves it on /guardrails (it joins the key group's rules with its own code) or declines it;
  • Test it runs everything that applies to you on pasted text; Draft turns a sentence into a rule as above.

The same list is /guardrails in the CLI. Signed in as an operator rather than an org member, the page says to sign in as a member or to set personal rules in venom.yaml (guardrails:).

Filters

/filters (the old /prompt-filters redirects here) holds rules that act on the text of prompts before they reach a daemon or provider, and on answers before they reach the member. The editor shows presets, the rule list, a rule editor with a live test box and the last 24 hours of hits, each linking to the log explorer.

Presets

Preset Rules
Secrets and passwords (secrets) redact API keys and tokens, private keys, cloud credentials, JWTs, passwords
Personal data (pii) redact card numbers, email addresses and phone numbers
Strict (strict) block secrets, private keys, cloud credentials, JWTs and passwords

Rule fields

Field Values Default Limit
name text none 80 characters
enabled on, off on
action redact, block, warn, log a redaction writes [REDACTED:<label>]
direction request, response, both request
match: keywords words, whole_word, case_sensitive, label 200 words, 100 characters each
match: regex pattern, case_sensitive, label 512 characters; runs on a linear-time engine, counted repeats up to 1000
match: detector credentials, secrets, passwords, private_keys, cloud_credentials, jwts, credit_cards, emails, phones, national_ids, prompt_injection
scope key groups, keys, models (empty means all) all 50 items each

Org-wide settings: filters on or off, which parts of a request they read (system messages, tool results, tool calls; all on by default), at most 50 rules, and an optional classifier model with inline or audit mode, a timeout (default 800 ms, 100 to 5000), fail open or closed, and up to 20 categories of up to 300 characters each.

Saving replaces the whole rule set and reports "Saved N rules. Every instance applies them within 10 seconds." The test box sends at most 64 KiB of pasted text with your unsaved draft, optionally as a response or with the classifier; it is never stored.

Secret scanner

On by default, the scanner reads every model answer passing through the cloud with the credentials detector.

Action Effect
redact (default) credentials in answers are replaced with a placeholder
block the answer stops at the credential
log hits are counted; the answer is unchanged

Changes are audited as prompt_filters.updated and secret_scanner.updated.

Troubleshooting

Message Cause and fix
"Choose an organisation in the org switcher…" you are signed in as an operator; pick an org first
"Your role cannot change prompt filters (org:manage)." only owners and admins change filters; developers can read and test
"Refused: the request did not come from this console." the form was posted from another origin; reload the page
"Your session ended. Sign in again." sign in again at /login
A request is refused with 403 prompt_blocked a filter rule with action block matched; the log explorer's filter field names the rule
A personal rule doesn't fire for a CI job personal rules apply only to your own keys, not to shared or key group keys