Blog
Security3 min read

Content rules you write as a sentence

Tell your firewall what your agents may not send or receive, in a plain sentence. Its AI checks every message against it, so reworded messages are caught too.

ControlClaw teamWaldos LLC

An agent that can send messages can send the wrong one. Someone in a big team asks it for the customer list, and it pastes it into a chat. Nobody meant harm, but the list is out.

Content rules let you say what your agents may not send or receive. Since 22 October you write a rule as a sentence, and your firewall's AI checks each message against it.

A rule is one sentence

Open Security rules, press New rule and pick Messages. Write what you want to keep out, for example "Don't ask about or share anyone's salary, pay or bonus."

A word list would miss most of the ways people say that. "How much does Anna make?" goes against the sentence. "When is the all-hands?" does not.

A content rule called Salaries and pay, set to Stop, with the Try a message box under it
A rule with its sentence, and Try a message under the list.

You can add exact words to the same rule, and detectors for passwords and keys, bank account numbers, card numbers, or five or more email addresses in one message. Four templates get you started: Salaries and pay, Customer data, Bank and card numbers, and Personal data in bulk.

Stop, hold or record

Each rule does one of three things when a message matches.

  • Stop. The message does not go. The agent is told why, and on Telegram and Slack the person in the chat gets a short note from the firewall.
  • Hold for approval. The people who get your confirmation codes receive the message's start, the rule and its reason, with two buttons, Send and Don't send. With no answer in 24 hours the message is dropped.
  • Only record. Nothing is stopped. The firewall notes that the rule matched.

Every template starts as Only record. Look at what it caught, then switch it to Stop or Hold.

Where it applies

The firewall checks messages on Telegram, Slack, email, website requests, webhooks and team chat. It also checks the voice in meetings and calls, and what agents do in the apps you connect, in both directions. WhatsApp is the one gap, because its messages are end-to-end encrypted on the agent.

One rule is always on and you can't turn it off. A message from an agent that looks like it contains a password, an API key, a private key or a token is stopped.

Try it before you save

Try a message sends a message you type to your firewall. It answers with the rule that would act and how sure the AI is, as a percentage. A rule acts from 50%. In our tests on 60 pairs of rules and messages, messages that broke a rule scored 63% to 99%, and plain ones 1% to 33%.

If a rule catches too much, write the sentence more narrowly.

Who sees the messages

The check runs on your firewall, which calls the model directly. ControlClaw's servers never see the message. The model is Jev, a decision model that answers yes or no with a score. It takes about 220 milliseconds a message, and costs about $0.04 per 1,000 messages from your included AI tokens.

Stopped and held texts stay on your firewall, encrypted, for seven days. Open them with What was stopped on Security rules. ControlClaw only learns which rule acted, where and when.

The firewall's list of messages held or stopped in the last seven days
What was stopped, on your firewall's own page.

If the AI can't answer, the rule acts anyway, so turning the AI off can't switch your rules off.

Two things to know

Every change to a content rule needs a 6-digit code that your firewall sends to you. That includes making a rule stricter. A rule's text reaches your agent and the AI, so ControlClaw's servers must not be able to set one alone.

Content rules catch mistakes and enforce policy. They are not a defence against someone who is trying to get around them on purpose. The firewall does not read what is inside files, photos, voice notes or attachments, so a picture of the text gets through.

Sentences need release 1.23.0 or later on your firewall. Press Update on the Firewall page if yours is older.

SecurityProduct

ControlClaw team

The people who build ControlClaw, a managed OpenClaw service with an AI firewall and approvals.

Keep reading

More from the blog

Agents3 min

Your agent in Google Meet

Your agent can join Google Meet, take notes, answer when you call it by name, and act on the meeting afterwards by instructions you write once.

Agents3 min

Start an agent from a template

Pick what an agent is for when you create it, and it is set up for you. Apply a template to a running agent, keep your own, or create several agents at once.