Guardrails are available on the Teams plan. Only owners and admins (or a custom role granted the right permission) can configure them. All workspace members are subject to guardrails once they’re turned on.
Open Guardrails settings
Click the account row at the bottom of the workspace sidebar, choose Workspace settings, then open Governance, then open Guardrails.Turn guardrails on
Toggle Guardrail protection on. If nothing has been configured yet, Genie starts from a set of protective defaults that you can adjust from there. At least one protection must stay active at all times — an empty guardrail configuration can’t be saved.Choose a provider
Guardrails are judged by Amazon Bedrock Guardrails (the default) or Google Model Armor. Pick one at the top of the page; both judge every message before Genie AI answers and show a blocked message the same way. Model Armor covers harmful-content strengths and manipulation attempts only. Off-limits topics, blocked words, the profanity filter and Check images stay in your settings but are marked as not supported while Model Armor is selected, and they don’t count as protection. Switch back at any time and your earlier settings apply again.What you can configure
Harmful content
Set a strength for each category — Off, Relaxed, Balanced, or Strict:Off-limits topics
Add named topics that Genie AI should refuse to discuss in your workspace — for example, “Investment advice.” Each topic gets a plain-language description of what counts as that topic, plus optional example messages to help Genie recognize it.Blocked words
Turn on a built-in profanity filter, and add your own list of banned words or phrases specific to your organization — one per line.Message shown when blocked
Optionally write your own message, shown to a member instead of Genie’s answer when a guardrail blocks their message. Leave it blank to use Genie’s default explanation.Scanning attachments
Guardrails judge a message’s text by default. Two more switches, both off until you turn them on, extend that to what’s attached:- Check images — attached PNG/JPEG images are judged for harmful content (hate, violence, sexual content, and so on) alongside the message’s text. Off-limits topics and blocked words only ever apply to text.
- Refuse attachments that cannot be checked — block a message outright if it carries something guardrails can’t judge at all: a PDF, a GIF or WebP image, or any image while Check images is off.
.txt, .docx, or spreadsheet is already covered as text — its content is extracted before Genie AI ever sees it.
Blocked messages log
The Guardrails page keeps a log of recent interventions: which category or topic triggered the block, where it happened (workspace chat or a project), who sent it, and when. The message content itself is never stored — only the reason it was blocked. Anyone who can already see this log can also turn on email alerts for new blocks — see Activity log.Manage guardrails from workspace chat
Ask the Genie workspace AI in plain language:- “Turn on guardrails and block sexual content and violence at the strictest level”
- “Add an off-limits topic for legal advice”
- “Show me recent blocked messages”
- “Turn off the manipulation-attempt filter”
Related pages
PII rules
Detect and block sensitive data like credit card numbers and API keys in chat.
Governance
Review projects and enforce compliance rules across your workspace.
Genie MCP Server
Manage guardrails programmatically from any MCP client.
Members and roles
Grant the guardrails permission to a custom role.