Arkose finds 97.7% of legitimate AI agent sessions carry no verified ID

Security firm Arkose has found that 97.7% of legitimate AI agent sessions carry no verifiable identity, PPC Land reported on October 21. In practice, the overwhelming majority of agents doing real work look indistinguishable from anonymous scripted traffic to the sites they touch.

The reason is structural. An agent is usually launched by a person and then acts on that person's behalf in a browser or in the cloud, but it carries no verifiable badge of its own: no signed credential in the request, no statement of which user it represents, what it is allowed to do, or when that permission expires.

That creates a bind for defenders. If a platform cannot separate an authorized agent from a disguised bot, it is left with two choices: block everything automated, or allow everything. The first breaks automation users explicitly asked for; the second opens the door to scraping, inventory hoarding and account abuse.

The pressure is moving from the margins into mainstream traffic. Retail, travel, ticketing and publishing sites are all seeing agents that compare prices, fill forms, hold seats and subscribe on a user's behalf. Risk systems built on IP frequency and device fingerprints misjudge traffic that has a real human behind it but behaves like automation.

The gap is in the authorization chain, not in detection. What a site needs is not an answer to “is this a robot” but proof of who authorized the agent, what scope was granted, and whether it can be revoked. Without that layer, the security cost of the agent economy is carried unilaterally by the platforms being visited.

Discussion of agent identity has picked up over the past year, drawing in payment and identity vendors, browser makers and security firms, with a rough consensus forming around agents carrying verifiable delegated credentials that can be refused when they exceed scope. The hard part is governance rather than cryptography: who issues the credential, who trusts it, and how it works across platforms.

Three things to watch: whether large platforms publish explicit policies for agent traffic, whether signed agent credentials reach real deployments, and whether the 97.7% figure moves. The speed at which it falls is roughly the speed at which agent identity gets built.