The Maze: Arkose Labs, which sells fraud and automated-traffic protection, says 97.7% of the AI-agent sessions it classified as legitimate lacked a verified product identity in July–September 2026. Its customer-network study raises a practical problem for retailers: software acting for shoppers can arrive without proving which service sent it.
Source: Arkose Labs. Original report chart compares signatures in Arkose-classified legitimate sessions with ChatGPT agent sessions.
The Details:
The missing credential identifies a product. A browser agent is software that opens websites and completes tasks for a person. Sites normally receive browser names and network information, but those signals are insufficient to prove the sender. Web Bot Auth adds a cryptographic signature: the agent signs a web request, and the receiving site checks it using a published verification key. Arkose found valid signatures in 2.3% of sessions it classified as legitimate. ChatGPT agent signed 78.4% of its own sessions. Those percentages describe different groups; the ChatGPT result cannot be applied to the entire sample. Nor does a verified product identity prove that a customer authorized a particular purchase.
The sample measures Arkose's view. Its report covers protected actions such as login, account registration and checkout. It counts identified agent products separately from other bots and automated indexing tools. Arkose classified 77.4% of agent sessions as legitimate; its malicious category means a session received an attack-risk classification. These are the vendor's judgments. It publishes neither raw session counts nor classification error rates. Technology, software-service and telecom customers supplied most agent traffic, and customer geography is undisclosed. The figures cannot establish the mix at a typical retailer or across the whole internet.
Wide distribution can coexist with little traffic. Named agents represented roughly one in 5,200 protected sessions during the quarter. Yet 65.2% of Arkose customers saw them on September 30, using a threshold of at least ten named-agent sessions on a protected action that day. One figure measures traffic volume; the other measures customer reach. The earlier reach baseline used four sampled days before personal-agent detection. September 30 included additional detections. Treating that comparison as pure adoption growth would confuse newly visible activity with newly occurring activity.
MarketMaze Take: A website needs two answers: which software sent this request, and whether this action should be allowed. A signature helps with the first. It does not settle the second. Arkose's useful contribution is the separation of those questions, rather than a universal fraud rate for shopping agents.
An unsigned visit is not a fraud verdict. The 97.7% headline concerns sessions already judged legitimate by Arkose. A retailer that blocks every unsigned agent could reject a real customer's task. The report does not measure how many orders such blocks lose, so the commercial cost remains a question for each merchant's own data.
Identification should narrow the uncertainty. For website operators anywhere, a verified sender can make an action easier to attribute while leaving purchase permission and abuse checks necessary. Arkose sells controls for this problem and has a stake in its framing. Its undisclosed sample size makes the findings a reason to examine actual customer traffic, rather than copy its percentages into a risk policy.
Sources: Arkose Labs, Cloudflare, PPC Land
