Meta's Muse sends most AI agent traffic and has no name to block

Quoted from HUMAN Security, State of Agentic Traffic, October 2026: In two months, Meta’s…
Quote: HUMAN Security, State of Agentic Traffic, October 2026

Summary

Meta's Muse made up 59% of the AI agent requests HUMAN Security saw in October, and it sends no bot name for a robots.txt line or a log filter to match.

Identify it by its Web Bot Auth signature instead: keep the Signature-Agent header in your logs and add the new muse.ai key directory to the ones your verifier trusts.

Meta’s Muse accounted for 59% of the AI agent requests reaching sites and apps protected by HUMAN Security in October, and Meta’s and OpenAI’s agents together made up about 90%, according to HUMAN’s October agentic traffic report. HUMAN sells bot and agent detection and publishes the report monthly.

Total agent requests in that data ran about 8 times August’s level. HUMAN counts one request as one action an agent takes on a site, such as loading a page or running a search, rather than a user or a session. Muse, the personal agent Meta launched on September 8, nearly quadrupled its request count in a month. It now sends 3.7 times as many requests as ChatGPT Agent, the next-largest agent in the report.

The agent with the largest share is also the one a log filter cannot pick out by name. Meta’s engineering post on Muse says that when the agent browses the internet, it “will appear as your activity”, and at launch Muse landed in logs as direct traffic from Linux Chrome. HUMAN itself identifies agents by combining behavioral signals, user-agent attribution and publisher-level integration rather than a string on its own. Robots.txt needs a user agent token to match, so grepping October’s logs for a Muse user agent will mostly come back empty.

What Muse does send, on some requests, is a cryptographic signature. Web Bot Auth, the same scheme Google uses to sign some of its agent requests, lets an automated client sign each HTTP request with a private key and add a header pointing at a public key directory on its own domain, which the receiving site can fetch to check the signature against. DataDome, which sells bot protection, reported on November 11 that Muse moved that directory from meta.com to muse.ai between November 7 and 9, so a verifier trusting a fixed list with meta.com on it now treats Muse’s requests as unverified. As of November 10, DataDome counted half of Muse’s requests arriving unsigned, with nothing to verify at all.

Where the requests land is the part that should decide policy, and HUMAN’s breakdown by destination puts almost everything in one bucket.

DestinationOctoberSeptember
Product and search85%81%
Content engagement4.3%3.6%
Account management3.2%4%
Authentication2.8%3.1%
Checkout and payment1.9%2.4%

HUMAN categorizes by the endpoint a request reached, not by what the agent was trying to do, so a hit on a checkout page does not mean an order was placed. Checkout’s share fell while its absolute volume grew 117% in a month, which makes the count rather than the percentage the thing to track. Agents, as the report puts it, “still browse far more than they buy”.

The mix of agents is also young enough that a list written a quarter ago now points at the wrong traffic. Perplexity’s Comet browser held 41% of this traffic in August and 3.6% in October, the only agent in the top 10 whose request volume shrank, while OpenAI’s Codex Browser nearly tripled its own. Both HUMAN and DataDome sell the detection these numbers come from, though DataDome’s own network also shows Muse as the most used agentic browser.

What to do

  • Confirm that your edge or log pipeline keeps the Signature-Agent header. Without it there is no signature to check and no way to separate a real Muse request from anything claiming to be one.
  • Add muse.ai to whatever directory list your verification uses, and treat that list as something that changes. Key directories get added and retired, and a fixed list fails quietly.
  • Decide what unsigned agent traffic gets before tightening anything. About half of Muse’s requests carry no signature, so a verify-then-allow rule leaves them in the same bucket as any unidentified visitor.
  • Set the rule by page type. Product and search pages take the overwhelming share of these requests, so the decisions worth arguing about are cart, checkout, login and account endpoints, not read access to a product listing.
  • Do not look for this traffic in analytics. HUMAN’s report says that on most sites analytics tools cannot separate AI agents from human visitors, so server logs and edge data are the only places it is visible.

HUMAN says its next report covers Black Friday and Cyber Monday, so the rules live this week are the ones that will run through the heaviest commerce traffic of the year.