Arkose: 97.7% of Legit AI Agent Sessions Lack Verified ID

Arkose Labs reported on Oct. 20 that 97.7% of legitimate AI agent sessions it tracked carried no verified identity signature.

If you run a high-ticket Shopify store, that number lands on your checkout. Your fraud filters and bot rules were built to sort humans from scripts, and a third group now sits between them: software acting for a real shopper without saying so. Block it and you may turn away a buyer who spent an hour researching a $2,400 order with an assistant. Wave it through and you accept sessions that look a lot like a fraud ring. I track shifts like this for high-ticket owners at Ecommerce Paradise because one bad rule costs real money at your order sizes.

Below is what Arkose measured and where its data is thin, how signed agents are supposed to work, what unsigned agents do to a store’s fraud logic, and five fixes you can make before Black Friday traffic arrives.

Lock down the store, then form your LLC with Northwest Registered Agent, the best LLC formation service for ecommerce owners. The reason it beats cheap formation sites that upsell you: Northwest does not sell your address and puts its own address on your public filings. Form your LLC with Northwest →

Arkose: 97.7% of Legitimate AI Agent Sessions Lack Verified ID

Arkose Labs published its State of Agentic AI Trust Report, Q3 2026 on Oct. 20, according to the Business Wire release. The report covers the vendor’s production traffic from July 1 to Sept. 30, 2026, across eight industries. It measures login, registration, checkout, password reset and one-time passcode flows, per PPC Land’s write-up.

The headline figure: 97.7% of sessions Arkose classed as legitimate AI agents had no verified Web Bot Auth signature. Only 2.3% did. Web Bot Auth is the method that lets an agent cryptographically sign its requests so a server can confirm who is knocking. ChatGPT’s agent was the exception, signing 78.4% of its own sessions, according to PPC Land.

Outside technology, disclosure was worse. PPC Land reports that 92.8% of agent sessions in non-tech sectors carried no disclosure at all.

Arkose supplied several more numbers, each with a baseline worth keeping in view:

Of all AI agent sessions, 77.4% were classified legitimate, against 7.9% of known bot sessions. By Sept. 30, 65.2% of Arkose’s customers saw agents on protected flows, up from a 50% baseline sampled over four days between July 15 and Sept. 16. About 1 in 5,200 protected sessions came from a named AI agent, roughly 0.02%. And 70.5% of registration sessions were fake-account attempts, a figure whose scope PPC Land says the report leaves unclear.

Two agents drew most of the attention. Meta’s Muse grew more than 30 times in sessions during September, per Business Wire, and Arkose’s detection for Muse went live Sept. 22. A second agent, Instinct, had 96.8% of its sessions look medium or high risk, yet neither Instinct nor Muse had a session linked to an attack. That finding rests on an eight-hour sample on Sept. 22 and 23 UTC, according to PPC Land.

Arkose sorts agents into three groups on its agentic AI page. The first is self-disclosing good agents that sign their requests. The second is non-disclosing good agents acting for a real user, a group PPC Land says includes Muse, Instinct and Perplexity’s Comet. The third is malicious agents run from cloud infrastructure for account takeover, fake accounts, payment fraud and scraping. The second group is the one that should worry a store owner, because it looks like a customer and behaves like automation.

Mitchell Davies, lead researcher on Arkose’s ACTIR team, framed the test this way: “The real question is, what did the agent come to do? What is its intent?” Kevin Gosschalk, Arkose’s founder and CEO, told Forbes that “blocking everything turns away deposits, and allowing everything is a blank check.” That comment was about banking, but Business Wire’s copy makes the same point for every sector: blanket allow-all or block-all will not work, and a retailer’s controls will differ from a bank’s.

Web Bot Auth and Agent Protocols: How Signed Agents Work

Cloudflare proposed Web Bot Auth in May 2025 as a replacement for user agent strings and IP addresses, both of which can be spoofed. On Oct. 24, 2025 it announced work with Visa and Mastercard built on it, per PPC Land’s coverage. Agents publish public keys in a directory and send HTTP Message Signatures. A tag in the Signature-Input header says whether the agent is browsing or paying, and Cloudflare said it would let site owners admit trusted agents registered with Visa or Mastercard through managed rulesets.

The standard is not finished. The IETF’s working draft, published Sept. 1, 2026 by Thibault Meunier of Cloudflare and Sandor Major of Google, is still an Internet-Draft with no RFC number, and it expires March 5, 2027, according to the IETF datatracker. Forbes reports that Visa, Mastercard and American Express have backed the Trusted Agent Protocol, which uses it.

Banks moved first on the defensive side. I covered how Citi and BofA restricted AI agents, joining JPMorgan, and Forbes quotes JPMorgan CFO Jeremy Barnum warning that customers could be liable if an agent misuses their credentials. On the card side, Amex introduced an intent check for AI agent purchases.

Retail is the open flank. Shopify opened its catalog to personal AI agents. Meta’s Muse can now check out with Shop Pay. The buyers are arriving before the identity layer is ready.

There is a real counterpoint to the Arkose numbers. Arkose sells the detection products it describes, and PPC Land notes the report is not independently audited and publishes percentages without raw counts. Agent traffic was concentrated among technology, SaaS and telco customers, so retail’s true share is unknown. The before-and-after comparison on reach is not like for like. Profound’s research says both Muse and Instinct use residential connections, which conflicts with Arkose’s account that only Instinct does, per PPC Land. Read this as a direction signal about the market, not a measurement of your store.

What Unsigned AI Agents Do to a High-Ticket Store’s Fraud Rules

My read is that the 97.7% matters less than the category it exposes. Your stack knows three things: customer, bot, fraudster. An agent shopping for a customer is a fourth thing, and at least one of them, Meta’s Muse, appears on networks as an ordinary Chrome browser, per Business Wire’s copy. Rules tuned to flag automated behavior will flag the shopper who is most ready to buy.

Shopify says AI-referred sessions convert 80% better, as I noted in my post on that claim. It is Shopify’s number and I would want to see your own, but if even half of it holds in your category, blocking agents is a revenue decision rather than a security setting. I also noted that Meta’s agent was told to rank known merchants first, which means trust signals will decide who gets shown.

The other side is just as real. If 70.5% of registration sessions across Arkose’s customers are fake-account attempts, open forms are being farmed. For a high-ticket store that shows up as junk accounts, junk leads in your CRM and email metrics you can no longer trust. Chargebacks are the expensive version, and I wrote up how to stop them in my guide to chargeback prevention for high-ticket stores.

Here is hypothetical math, not reported data. Say your store books 300 orders a month at an $1,800 average, which is $540,000. If 1% of orders start with an agent session and your rules block them, you lose 3 orders and $5,400 a month. If 1% of orders are fraudulent and you wave everything through, you take the same 3 orders as chargebacks, plus dispute fees and the product cost. At these order sizes both mistakes hurt, which is why a blunt rule in either direction is the wrong answer.

My rough thresholds, which are mine and not a standard: under 1% of checkout sessions from suspected agents, tag them and watch. Between 1% and 5%, write explicit allow and challenge rules. Above 5%, treat agents as a channel with its own reporting, the way you would treat a new marketplace.

What I’d do with signed versus unsigned: reward signed agents, because a verified signature is the strongest trust signal available right now. Do not block unsigned agents by default. Add friction where money moves, which means login, registration and any order above a value you pick. Screening services such as ClearSale exist for this kind of order review, and a phone call to a $3,000 buyer still beats any filter. High-ticket customers expect that call anyway.

The wider point is that the checkout you built two years ago assumed a human at a keyboard. Google’s Universal Cart is live, and Newegg has turned on Google checkout in Gemini and AI Mode. Your store on Shopify inherits whatever defaults you never changed.

Most owners will not have time to audit all of that between now and January. If you would rather hand the build and the rules to people who do this daily, that is what my turnkey done-for-you service is for.

Not sure how much agent traffic your store gets? Compare numbers with other high-ticket owners and me inside the community. Join the Skool community →

Five Fixes for Agent Traffic on Your Checkout Before Q4

Five moves, in order:

Measure first. Pull 30 days of checkout and registration sessions from your analytics and your CDN or firewall dashboard, then count how many show automation signals yet still ended in a normal paid order. That tells you whether you sit near Arkose’s 0.02% average or closer to 5%. If you want a second set of eyes on the numbers, book a discovery call.
Reward signed agents and challenge unsigned ones. If your traffic runs through Cloudflare, check whether its managed rules for verified agents are available on your plan, since the company said in October 2025 it would offer them. Then add a challenge, not a block, on login, registration and checkout. My notes on the AI-agent readiness audit show where most Shopify stores fail today.
Put a review step on big orders. Pick a threshold, for example every order over $1,500, and route it to manual review or a screening tool before it ships. Call the buyer. Check that the shipping address is not a forwarder you have never seen.
Log evidence for disputes. Visa widens its chargeback evidence rules on Oct. 24, as covered in my post on the change. Store the IP, timestamps, signature status and agent name with each order so you can answer a dispute with facts. An accounting tool such as Finaloop keeps the books side clean while you do it.
Clean up registration and email. Turn on double opt-in in Klaviyo, tighten signup forms, and audit the forms feeding your CRM. If you run HubSpot, my setup guide for the free tier covers the configuration.

Run the Black Friday version of this too. My high-ticket Black Friday stress test pairs well with the five moves above. My notes on security infrastructure for scaling a store cover the basics underneath.

If your processor is a single point of failure, fix that before the rush. I laid out a backup plan for payment processors. A second rail through Airwallex is one option worth a look.

Frequently Asked Questions

What did Arkose actually find?
Across July 1 to Sept. 30, 2026, 97.7% of sessions it classed as legitimate AI agents had no verified Web Bot Auth signature, per PPC Land. Arkose is a vendor, so treat the figures as directional.

Are AI shopping agents safe to allow at checkout?
There is not enough evidence to say. Arkose saw no attacks linked to Muse or Instinct, but that finding rests on an eight-hour sample. Allow them with friction on high-value orders, as covered in my post on the Shop Pay AI checkout security bug.

What is Web Bot Auth?
It is a method for agents to sign HTTP requests with a cryptographic key so servers can verify them. It is still an IETF working draft, according to the datatracker, so expect changes.

Should I block AI agents on my Shopify store?
Not by default. Blocking everything risks losing buyers, and Shopify says AI-referred sessions convert better, though that is its own claim. Challenge unsigned agents at login and checkout instead.

How many agent sessions will my store see?
Arkose’s average is about 1 in 5,200 protected sessions, but its customers skew toward tech, SaaS and telco. Measure your own for 30 days. The saved card data lock-in numbers I covered show how little control most firms have over payment data in general.

I do not have a store yet. Where do I start?
Grab my free high-ticket niches list. Then read my guide on what high-ticket dropshipping is. Pick the niche first and worry about agent rules once orders start.

Want my team to build and run your high-ticket store while you watch the agent traffic numbers? See the turnkey done-for-you service →

I will keep tracking the agent identity standards as they firm up, and I will flag it when a processor or platform starts requiring signatures. Subscribe to the YouTube channel for daily breakdowns. More breaking news coming through the day.

If this was useful, these go deeper:

Trevor Fenner

Trevor Fenner is an ecommerce entrepreneur and the founder of Ecommerce Paradise, a platform focused on helping entrepreneurs build and scale profitable high-ticket ecommerce and dropshipping businesses. With over a decade of hands-on experience, Trevor specializes in high-ticket dropshipping strategy, niche and product selection, supplier recruiting and onboarding, Google & Bing Shopping ads, ecommerce SEO, and systems-driven automation and scaling. Through Ecommerce Paradise, he provides free education via in-depth guides like How to Start High-Ticket Dropshipping, advanced training through the High-Ticket Dropshipping Masterclass, and fully done-for-you turnkey ecommerce services for entrepreneurs who want a faster, more hands-off path to growth. Trevor is known for emphasizing sustainable, real-world ecommerce models over hype-driven tactics, helping store owners build scalable, sellable, and location-independent brands.