Arkose Labs data shows that legitimate AI agents and malicious automation can look alike, leaving websites to distinguish customer activity from fraud.

A single hacker's AI agents probed more than 1,200 websites and reached sites belonging to more than 100 companies, among them Deere, SharkNinja and Cummins, according to a Reuters exclusive published Oct. 19. The operator ran the campaign on Cairn, an open-source agent framework that NorForge describes as a general-purpose problem-solving engine, one that works in stages from a starting point toward a defined goal. Snorre Fagerland, principal threat researcher at NorForge, told Reuters the campaign used roughly $5,000 in tokens, or about $50 per successful break-in. NorForge's own report said the firm had no actual token-usage data, estimated an average of around $5 per task once unsuccessful and early-ending projects are counted, and warned that the estimate could be well off. Neither Reuters nor The Hacker News, which covered the findings the next day, could say whether any data was taken.

The agents' main objective was code execution on web servers, though their task instructions also listed login SQL injection and account takeover as priorities. Spotting an automated session does not reveal whether it is attacking a site or acting for a customer. A day after the Reuters report, Arkose Labs published data on the account flows where customers log in, sign up and pay, and on the agents that arrive there on behalf of real people. In its Q3 data, 77.4% of AI agent sessions were classified as legitimate, compared with 7.9% of known bot sessions.

About the data. Arkose Labs production traffic on login, registration, checkout, password reset and one-time passcode flows, July 1 to Sept. 30, 2026. The report publishes percentages and multiples, not raw counts, and most agent sessions came from technology, SaaS and telco customers. Muse figures before Sept. 22, when live detection began, were reconstructed from stored fingerprint data, and the Muse and Instinct technical profiles cover the first eight hours of detection. Arkose Labs did not observe the Cairn campaign.

How many legitimate AI agents are already on login and checkout pages?

The Q3 2026 edition of The State of Agentic AI Trust Report found AI agents on login and registration flows in all eight industries Arkose Labs measured, from banking and travel to gaming and streaming. By Sept. 30, 65.2% of its customers recorded at least 10 named AI agent sessions in a day on at least one protected flow, against 50% on sampled days before Arkose Labs began detecting personal AI agents. The Sept. 30 count includes personal agents and Grok Bot and the earlier days do not, so the two figures show reach under expanded detection rather than a like-for-like growth rate.

Password reset showed the widest spread. On the earlier sampled days, 40% of customers saw agents on that flow, and on Sept. 30, 77.8% did. Support forms moved from 25% to 50% on the same basis.

Volume is still modest. Roughly one in every 5,200 protected sessions came from a named AI agent, though the report stressed that this small share already reaches most of the customer base.

Can websites verify which AI agents are legitimate?

Verified signatures remain rare. Only 2.3% of legitimate agent sessions presented a verified Web Bot Auth signature, which lets a site confirm cryptographically which agent product sent a request. The remaining 97.7% carried no disclosure. Cloudflare introduced the approach in May 2025 to verify bots beyond user-agent headers, which it described as easily spoofed, and published IP ranges, which it described as brittle. Seventeen months later, adoption among the agents Arkose Labs tracks remains thin.

CHART 1: Arkose Labs chart "Almost no legitimate agent sessions carry a verified signature" (Key finding 1). Verified 2.3% vs. no disclosure 97.7%, with ChatGPT agent 78.4% callout. Source: Arkose Labs, Q3 2026, protected actions only.

ChatGPT agent signed 78.4% of its sessions and is the only agent the report found signing at scale. Grok Bot for Enterprise signs its requests as well. The network-wide rate still held at 2.3%, and among customers outside technology, 92.8% of agent sessions carried no disclosure.

Arkose Labs groups agent traffic into three populations. Population One covers good agents that disclose themselves. Population Two covers good agents that act for a real user without declaring themselves, including personal agents such as Instinct and Muse and agentic browsers such as Perplexity's Comet running on a person's own laptop. Population Three covers malicious agents running account takeover, fake account creation and payment fraud, usually from the cloud and at scale. Legitimate agent traffic in Q3 fell almost entirely into Population Two. Telling good AI agents from bad ones therefore depends on spotting agents that never announce themselves, since a site that waits for a signature will miss nearly every one acting for a customer.

Kevin Gosschalk, Arkose Labs founder and CEO, said in the company's Oct. 20 announcement of the report that the CISO of one of the world's largest banks told a recent banking industry summit that, for the first time in his career, a technology has arrived ahead of the controls.

Is bot detection enough for AI agents?

The report argues that automation signals can flag a session as automated without revealing its purpose. Synthetic browsers, cloud infrastructure and machine-timed clicks are common to both groups, so a rule keyed to those traits would sweep up many agents working for real customers.

CHART 2: Arkose Labs chart "Sessions classified as legitimate" (Key finding 2). AI agent sessions 77.4% vs. known bot sessions 7.9%. Source: Arkose Labs, Q3 2026.

Instinct, a personal agent, shows how far appearance can mislead. It connects over home and mobile networks and cycles through device fingerprints, a method attackers favor for slipping past reputation checks. In the first eight hours of live detection, Arkose Labs rated 96.8% of Instinct sessions medium or high risk, yet none was linked to an attack, so a policy that blocked on risk rating alone would have turned away nearly all of them. Muse, by comparison, was rated medium or high risk in 24.9% of sessions and likewise had none linked to an attack.

CHART 3: Arkose Labs chart "Rated risky vs. linked to an attack" (Key finding 5). Instinct listed first: rated medium/high risk 96.8%, linked to an attack 0%; Muse 24.9% / 0%. Source: Arkose Labs, first eight hours after detection went live, Sept. 22–23 UTC.

The two agents are built in opposite ways. Just 0.2% of Instinct sessions came from hosting or data-center IPs, compared with 4.3% for people on the same sites, while Muse operates from the cloud on a small, stable set of fingerprints and sent 59% of its sessions from data-center IPs.

CHART 4: Arkose Labs chart "Sessions from hosting or data-center IPs" (Key finding 5). Muse 59%, people 4.3%, Instinct 0.2%. Source: Arkose Labs, first eight hours after detection went live, Sept. 22–23 UTC. Mitchell Davies, lead researcher on the Arkose Cyber Threat Intelligence Research (ACTIR) team, in the same announcement. "The real question is, what did the agent come to do? What is its intent?"

How much fraud sits at the same doors?

The same flows draw heavy attack traffic. These rates cover every session at protected endpoints, agent or otherwise. In Q3, 70.5% of registration sessions were fake-account attempts, 33.5% of sign-in sessions were classified as illegitimate and 17.7% of checkout sessions were card testing or payment fraud.

With fraud at those levels, waving all automated traffic through would hand attackers the registration and checkout flows, and the report said blocking all of it would cost businesses the real customers and revenue that arrive through legitimate agents.

What comes next for agent traffic?

Disclosure is spreading slowly. Grok Bot for Enterprise signs its requests, and Sigma switched during the quarter from an undeclared agentic browser to one that identifies itself. Google's Gemini Spark, still in beta, does not identify itself, and neither do the personal agents driving the most growth. Muse sessions grew more than 30-fold in September, the report said, while Perplexity Comet, ChatGPT agent, Grok and Gemini Spark held roughly flat.

Arkose Labs expects agents to move beyond the browser as soon as Q4, operating a person's own phone or running Android inside virtual machines of their own. An agent working from a real handset gives a site no virtual machine to fingerprint, and mobile agent traffic on protected flows was negligible in Q3. The report also said Muse is set to add end-to-end encryption, which would keep even Meta from seeing what users ask it to do. Once neither the site nor the agent's developer can read the instructions, the report argued, what the agent does is the only remaining signal of intent.

How can you allow good AI agents without blocking customers?

The report makes four recommendations. Stop treating "bot" as a verdict, since automation signals describe what a session is. Welcome agents that sign with Web Bot Auth, and judge every session on its behavior whether it is signed or not. Measure agent traffic on every protected flow, including sign-up, sign-in, checkout and support. Plan for the agent mix to keep shifting, because a new agent can become a leading source of traffic within weeks.

That approach is the core of agent trust management, in which each session is classified by what it attempts and a policy is applied to that action, so a business can admit agents acting for customers while keeping its controls against account abuse and fraud. For retailers, banks and gaming operators preparing for agentic commerce, the first step is visibility into which agents reach which flows. Arkose Labs offers Agent Trust Manager, part of its Arkose Titan platform, to classify sessions and set policy by agent type and endpoint.

The full Q3 report is available now, and Arkose Labs' webinar on personal AI agents is available on demand. The company said the Q3 report is the first in a quarterly series, and the Q4 edition, due in early 2027, will compare quarters on how agent traffic grows and how often agents identify themselves.

The Hacker News https://blogger.googleusercontent.com/img/b/R29vZ2xl/AVvXsEgQQyjwPYjJP0wddSEB8Dlpr3dlnQUs52-WmlrZfqJoBPeOvv2Zoqlq-FhEAz_Xeprj_mtrI1MGCW1JS840JUjVEK6VoNe6zCNNTw_7YmyvNmf3E5pprZ3zqP8lszq74Wt97SvbJo5yeuyep0U6-nGs0vdarg4_WUrc5r6L0ML0xE-BsPipJd2-1PMHTvO1/s76-e365/thn.jpg
Found this article interesting? This article is a contributed piece from one of our valued partners. Follow us on Twitter and LinkedIn to read more exclusive content we post.