On 2026-10-18 our own outreach agent sent nine emails, and every one of them ended with a signature that said "[REDACTED]" where the founder's email address belonged. The agent had fetched the right signature from Signatoro, the agentic email signature platform, before each send. The signature broke afterwards, on its way from the fetch to the send, because the agent typed it out itself. The fix: the agent's code puts the signature into the message, and the agent reads each sent email back.
What the emails looked like
The address was missing twice in each email:
- In the plain-text part, the contact line read "[REDACTED]".
- In the HTML part, the link text read "[REDACTED]" and the link pointed to
mailto:%5BREDACTED%5D, which is "[REDACTED]" written as a URL.
Everything else in the signature was intact: the name, the title and the links.
What went wrong
The agent ran each send in two steps:
- It called
get_signatureand got the signature back. The reply had the real address, and so did the files the agent built the email from. - It sent the email by calling its Gmail tool, typing the subject, the body and the HTML body into that tool call.
Step 2 is where the address was lost. The agent's context masks email addresses automatically, so when the model wrote the HTML body out again, the address came out as "[REDACTED]", and that is what Gmail received. Nothing read the sent message afterwards, so all nine went out before anyone saw it.
Signatoro returned the correct signature every time. The failure was in the copy: the signature passed through the model a second time, and the model's context changed it.
Why this can happen to any agent
An agent that fetches a signature and then retypes it into a send call puts the signature through the model a second time. Whatever the model's context does to text on the way, such as masking email addresses, happens to the signature too. The signature it fetched was right, and the one it sent was not.
The same goes for any detail the model copies by hand. The safest signature is one the model never types.
How to send the signature intact
- Insert the signature with code. Build the message in code from the API response, or from a file the agent saved it to, so
htmlandtextreach the mail API byte for byte. One script that fetches the signature and sends the email in the same step is simplest. - Never retype it into a tool call. If the only way to send is a tool that takes the body as an argument, pass a file or a variable, not text the model writes out again.
- Read each sent message back. Before the next send, open the sent email and check the address, the phone and the
mailto:link are still there. A search for "[REDACTED]" catches this exact failure.
Signatoro's get_signature reply, from the API and the MCP server, now carries this rule in its usage note, so an agent that never opens the guide still reads it. The agent guide has the details.
Let the sending tool add it
The most reliable setup takes the signature out of the model's hands completely: the tool that sends the email adds it at send time. That is how Signatoro's community node for self-hosted n8n works. The model writes the body, the node fetches the saved signature and puts it after the body, and the send step sends the result. The model never sees the signature, so it can't change it.
The same holds for any workflow tool or script: the model writes the words, and code adds the signature.
Questions
Why does my AI agent's email signature say [REDACTED]?
The agent copied the signature by typing it out, and something on the way masked the email address: some agent contexts mask email addresses automatically. Have the agent's code insert the signature from the API response or a file instead.
Is the signature wrong in Signatoro?
No. In our case the API returned the real address every time; the address was lost after the fetch. Check the signature on your signature page: if it shows the address there, the problem is in how the agent sends.
How do I catch it before the next email goes out?
Read the sent message back after each send and look for "[REDACTED]", a missing address or phone, or a mailto: link that doesn't point to a real address. Stop the run if one fails.
Related: Stop your AI assistant from inventing your phone number, title or address.