Are a model provider's safety filters enough?

A model provider's safety filters stop content the provider forbids for everyone. They do not know your product's rules, your data or your agent's tools, which is what separate guardrails are for.

Side by side

Provider safety filtersYour own guardrails
Whose rulesThe provider's usage policyYours
CoversBroad harm categoriesAny policy you write
Knows your dataNoYes, through your policies
Tool callsNot judged against your rulesChecked before they run
Across providersDifferent for each modelThe same everywhere
EvidenceLittleA log of each verdict and why

What the provider's filters are good for

They are free, already on, and catch the worst general content. Keep them.

What they miss

  • Your rules. "Do not quote another customer's order" or "do not discuss competitors" is no provider's concern.
  • Actions. Whether a tool call is safe depends on your systems.
  • Consistency. Change model or provider and the behaviour changes with it. Open-weight models you host come with no filter at all.
  • Tuning. You cannot adjust a provider's threshold when it blocks your honest users.
  • Audit. A refusal with no reason is hard to show a regulator.

The usual setup

Provider filters as the floor, your own guardrails for everything specific to you. See system prompt vs guardrails for why instructions to the model are not a third layer you can rely on.

With White Circle

Control works with any model or setup, through an API or as middleware, so the same policies hold when the model behind them changes.

Train your own models

Post-train any Hugging Face model with Halo. Up to 4× faster than TRL