Arkose Labs, a bot-detection vendor, published its State of Agentic AI Trust Report for the third quarter of 2026 on October 20. Its central figure: only 2.3% of sessions by legitimate AI agents carried a verified Web Bot Auth signature between July 1 and September 30, so 97.7% did not. As agents move from novelty to routine traffic on login, registration and password-reset pages, the report is a measurement of how little of that traffic can currently prove who it is. The data is Arkose's own and unaudited, and the company sells products in this space, which shapes how the findings should be read.
OpenAI
Grok
PerplexityWhat was measured
The report draws on Arkose production traffic from July 1 to September 30, 2026, excluding test, internal and analyst accounts, across eight anonymised industries, with agent traffic concentrated in technology, SaaS and telecoms. It covers five protected actions: login, registration, checkout, password reset and one-time passcode. Arkose publishes only percentages. It states that no raw counts are published, so neither the total number of sessions nor the number of customers can be assessed.
Web Bot Auth is a mechanism by which an automated client signs its requests so a site can verify the operator. A signed agent can be allowed through with confidence; an unsigned one has to be judged by behaviour, network and fingerprint signals. The headline gap, then, is between agents that identify themselves and agents that merely look legitimate to a classifier.
Most legitimate agents are unsigned
Arkose classifies 77.4% of AI agent sessions as legitimate, against 7.9% of known bot sessions. It divides traffic into three groups: self-disclosing good agents, non-disclosing good agents, and malicious agents. Nearly all legitimate traffic sat in the second group. Among signers, ChatGPT agent accounted for 78.4% of signed sessions, and outside technology companies 92.8% of agent sessions carried no disclosure. The report does not explain why technology customers, who carry most agent traffic, show lower disclosure.
The classification of "legitimate" is Arkose's own, and its criteria are not fully described in the coverage reviewed. That is the largest limitation: the 2.3% figure is a share of sessions Arkose already judged to be legitimate, so the unsigned majority is, by construction, a population that identity checks would otherwise have to clear another way.
The gap is not between good and bad agents. It is between agents that can prove who they are and agents that only look as if they should be let in.
Analysis of Arkose Labs' Q3 2026 findings
Where agents are showing up
Share of customers seeing AI agents on protected flows
Percent of Arkose customers with at least 10 named-agent sessions on a flow in a day, baseline vs September 30.
Source: Arkose Labs, State of Agentic AI Trust Report Q3 2026. Detection for personal agents began mid-window, so the two columns are not like for like.
The share of customers seeing agents on at least one protected flow rose from a 50% baseline to 65.2% on September 30. The largest moves were on password reset (40% to 77.8%), registration (34.9% to 55.6%) and support forms (25% to 50%), while login rose from 62.1% to 74.2%. Checkout moved from 35% to 40%, and OTP and SMS flows slipped from 43.1% to 40%. A customer counted only if it had at least 10 named agent sessions on a flow in a day.
One caveat is built into the data. Detection for personal agents began during the window, so the baseline, an average of four sampled days between July 15 and September 16, was measured before that detection was active. The before and after columns are therefore not a like-for-like comparison, and part of the increase is likely improved detection rather than purely more agents. Named agents were roughly 1 in 5,200 protected sessions, about 0.02%, so in absolute terms this is still a thin slice of traffic. Separately, 70.5% of registration sessions were fake-account attempts, though the report does not say whether that covers all registration sessions or only automated ones.
New agents: Muse and Instinct
The report singles out two newer agents. Meta's Muse launched on September 8, and Arkose's detection for it went live on September 22. Muse sessions grew more than 30 times in September, with earlier figures reconstructed from stored fingerprint data, and live and reconstructed detection matched exactly from September 26. Amazon blocked Muse on September 20. Instinct peaked around September 19, and in the first eight hours after detection went live, 96.8% of its sessions were rated medium or high risk. Neither agent had a session linked to an attack in that sample.
Those profiles come from a narrow window, and Arkose describes Muse as cloud-hosted and Instinct as using home and mobile networks with rotating fingerprints, while another research firm's work says both route through residential connections, which conflicts with the report. Across established agents, volume grew 37.8% from July to September, while Perplexity Comet fell 30.1% over the same period. Claude Desktop traffic was excluded because its detection also counts people logging into Claude.
Why it matters
For site operators, the trade-off is concrete. If agents cannot be verified, operators have two blunt options: block automation that includes legitimate customer-authorised tasks, or admit it and rely on behavioural signals that are easy to evade. For agent builders, the options are to sign requests, which exposes the operator to accountability, or to remain anonymous, which risks being blocked. The 97.7% figure suggests adoption of signing is at an early stage. Enterprises that run agents across many providers typically want a single point of control for identity, logging and permissions, a design goal shared by gateways and workspaces such as Metir.
What to watch next
- Whether more agent vendors adopt Web Bot Auth signing, and whether the signed share rises materially in Q4.
- Whether Arkose or other vendors publish raw counts and criteria for "legitimate".
- How retailers and banks respond to agents on checkout and password-reset flows.
Sources:
- PPC Land: 97.7% of legitimate AI agent sessions carry no verified ID, Arkose finds (October 20, 2026)
- AI Weekly: AI News Today, October 21, 2026
Image credits
Header and in-body image: Dell server racks in a data center, via Wikimedia Commons, licensed under CC BY-SA 3.0. It is illustrative and not a facility named in the report.