The name on the request was the credential

Cequence and KPMG published field data on 11 November 2026 in which traffic claiming one consumer AI agent’s name hit a single marketplace 414,210 times in a week — more than every genuine agent visit the same vendor counted across five industries. Every one of those visits went to the sign-in page and every one failed, which is credential stuffing wearing an allowlisted name. The lesson is not that the allowlist was too generous: it is that an unauthenticated string was being used as a credential, and the control that caught this was behavioural, not an identity check.

By Agentic AI Wiki 10 min read

Somewhere in your stack is a rule that says automated traffic is blocked unless it is one of the good agents, and it identifies the good agents by reading a string the client typed. On 11 November 2026 Cequence Security and KPMG published the bill for that: 414,210 visits to a single online marketplace in one week claiming a consumer AI agent's name, every one of them aimed at the sign-in page, every one of them failing. The name was the credential, and credentials that cost nothing to mint get minted.

What the report counted

The source is Agentic AI Runtime Security, Governance, and Control, published on 11 November 2026 by Cequence Security with input from KPMG Singapore, and built on traffic Cequence sees across its own customers. Four numbers carry the argument.

MeasurementFigureWindowReading
Agent visits counted across five industries 281,627 15–21 Oct 2026 46% reads, 39% shopping — the baseline of real agent traffic.
Share of those that took an action ~1 in 7 same week Booked, bought, signed in or managed an account for a customer.
Visits spoofing one agent at one marketplace 414,210 one week More than every genuine agent visit above. All sign-in, all failed.
Unidentified agent traffic at a Tier 1 carrier ~100/day → 32,000/day three weeks Growth after one consumer agent launched — a capacity event too.
Spoofed against genuine agent visits, October 2026 Three measurements from the Cequence and KPMG report. Traffic claiming one consumer AI agent's name at a single online marketplace reached 414,210 visits in a week. Genuine agent visits counted across five industries over 15 to 21 October came to 281,627. Of those genuine visits, about one in seven, roughly 40,000, went on to book, buy, sign in or manage an account. Agent visits counted, October 2026 Spoofed, one marketplace 414,210 Genuine, five industries 281,627 Of those, acted on account ~40,000 0 100k 200k 300k 400k One week of spoofed traffic at a single site outweighed every genuine agent visit counted across five industries. Single-vendor figures, one week each. The spoofed total is not a like-for-like sample of the genuine one.
One week of forged identity at one site outweighed the genuine agent traffic the same vendor counted across five industries.

Two provenance notes before anyone quotes these. These are single-vendor figures over single weeks, and the 414,210 total is from one marketplace while the 281,627 is a five-industry sample — they are not a like-for-like ratio, so the honest statement is that forged traffic at one site exceeded genuine traffic across the sample, not that forgery outnumbers legitimate use on the web. And the impersonated name: the press materials identify it as Google's agent and state that Google was not involved, while some coverage describes it only as a widely used consumer agent. Which brand was borrowed does not change the mechanism, and the mechanism is the story.

A User-Agent string was always a credential. It only now has a price.

Nothing here is a new vulnerability. HTTP has never authenticated the User-Agent header, every engineer knows it, and for twenty years it did not matter much because the header bought you nothing. What changed in 2026 is that sites started attaching value to specific values of it.

Follow the incentive. Publishers and retailers decided that agent traffic was worth serving rather than blocking, so they built carve-outs: higher rate limits for named agents, structured responses, exemption from the challenge that a human visitor would get, sometimes a route that skips the bot-management layer entirely. Each carve-out is a privilege. Each privilege is addressed to a name. And the name is self-asserted, which makes the whole arrangement a bearer credential with the secret printed on the outside.

So the correct reading of 414,210 sign-in attempts is not that someone discovered a clever bypass. It is that someone read the access policy as documented and complied with it. The attacker's tooling did exactly what the policy asked — present as an approved agent — and the policy had no step that could tell the difference. This is the same confusion that ambient authority describes from the other direction: a permission attached to an environment rather than to a verified action ends up held by whoever can stand in that environment.

There is a second, duller finding in the same body of work that makes the point harder to dismiss as one bad week. Cequence separately describes an actor impersonating 32 different bots — crawlers and assistants attributed to OpenAI, Anthropic, Perplexity, Meta and others — while hunting for credentials and keys. Thirty-two names is not a targeted attack on one brand's reputation. It is an enumeration of an allowlist.

What a signature fixes, and the two things it does not

Three-bucket agent access policy An incoming agent request passes a signature verification step and lands in one of three buckets: signed with a recognised operator, which earns rate-limit carve-outs; signed with an unrecognised operator, which is logged as its own class; and unsigned, still the majority, where behaviour decides. Behavioural controls apply across all three buckets. Incoming request Verify RFC 9421 Signed · operator known carve-outs Signed · unrecognised log as own class Unsigned — the majority behaviour decides Behavioural controls still decide every bucket
Verification turns a two-way guess into a three-way decision. It does not remove the third bucket, which is still the biggest.

The fix for a forgeable name is an unforgeable one, and it exists. Web Bot Auth, adopted as the IETF webbotauth working group's first document in September 2026 and built on RFC 9421 HTTP Message Signatures, has an agent sign its requests with a key whose public half its operator publishes. Cloudflare folded message-signature verification into its Verified Bots Program; Akamai, AWS WAF, HUMAN Security, Vercel and Stytch are also reported as verifiers. Against impersonation this is close to total, because the claim stops being a string.

Then two things stay broken, and both matter more than the part that got fixed.

Signing adoption is far behind verifying adoption

The signers that recur in reporting are Google, OpenAI and Amazon Bedrock AgentCore. That is a short list against the population of automation that will actually reach your origin. A rule of the form signed or blocked is therefore, as of late 2026, a rule that blocks most legitimate agents along with the forgeries — which is why the useful deployment is three buckets rather than two, and why the unsigned bucket cannot default to deny yet.

A signature names an operator, never an errand

This is the one that gets mis-specified in design documents. A valid signature establishes that the request came from software run by a particular company. It does not establish which of that company's millions of users is behind it, whether that user asked for this, or what they authorised — the signature is byte-identical whoever the person is, because the key belongs to the operator. So a signed agent carrying a prompt injection verifies perfectly, and abuse by one of an operator's users looks exactly like use by any other. Binding a request to a person is a delegation problem, answered by OAuth-style grants, by Poppy for a third party's agent signing in as your customer, or by a mandate credential as in AP2. Web Bot Auth sits underneath all of them and substitutes for none.

What actually caught it was shape, not identity

Worth noticing how this detection worked, because it did not involve verifying anybody. The spoofed traffic was identifiable because of what it did: it only ever touched the sign-in route, and it failed every single time. Cequence's own framing of the rule is about as blunt as security advice gets — if something only ever visits the sign-in page and fails every attempt, that is abuse whatever name it gives itself.

That discriminator has a property no identity check has: it is indifferent to the claim. It would have caught this traffic if it had presented as Chrome, as a signed agent from a real operator, or as nothing at all. And the inverse holds too — a perfectly signed agent from a reputable operator, doing the same thing, is also abuse, and your signature verifier returns valid for it.

The practical ordering follows. Behavioural controls are the load-bearing layer and signatures are an input that makes one class of lie expensive. Deploying it the other way round — treating verification as the control and keeping behaviour as a legacy fallback — reproduces the original mistake at higher cost, because you will have built a carve-out for verified traffic and will have spent the budget that used to watch what traffic did.

What to change this week

Three things, in order, none of them large.

Find every place a name grants something. Grep your edge configuration, WAF rules and application middleware for user-agent matching, and list what each match buys: a rate limit, a cache behaviour, a skipped challenge, a route. That list is your exposure, and most teams are surprised by its length because the entries were added one incident at a time.

Split the carve-outs from the identification. Keep the carve-outs — serving agents well is a reasonable business decision, and the retrieval-quality consequences of refusing them fall on your own content. But move each one behind a check that is not a string: a signature where the operator signs, a first-party key or account where you have a relationship, and otherwise the behavioural layer. The carve-out that cannot be moved behind any of those is a carve-out you should not have.

Instrument the sign-in route by claimed identity. The specific signal in this report is cheap to reproduce: group authentication attempts by claimed agent name and alert on a group whose failure rate approaches 100% or whose only touched route is the login endpoint. You are looking for a population that never succeeds, which is not a threshold anyone has to tune.

And one thing not to do. Do not read this as an argument for blocking agents. The same report finds roughly one in seven genuine agent visits taking a real action for a customer — booking, buying, signing in, managing an account — which is a channel, not noise. The failure here was not hospitality. It was issuing a credential that anyone could type.

FAQ

Is spoofing an AI agent's User-Agent a new attack?

No. Forging the header is as old as HTTP and requires no skill. What is new is that forging it now pays, because sites began granting named agents privileges that human visitors do not get.

Would Web Bot Auth have stopped this specific attack?

Yes, if the site had required signatures on the privileged path — the forged traffic could not have produced a valid signature for someone else's key. But requiring signatures today also blocks the many legitimate agents that do not yet sign, which is the trade the three-bucket policy exists to manage.

Does a verified signature mean the request is authorised?

No, and this is the most common design error. It means the request came from a particular operator's software. It says nothing about which user is behind it or what they approved, so it must never be wired into the place where a user identity or a permission is checked.

Was Google responsible for the traffic claiming to be its agent?

No. The press materials naming Google's agent state explicitly that Google was not involved. That is the whole point of impersonation: the operator whose name is borrowed is a victim of it, not a participant.

How do I tell a credential-stuffing bot from a real shopping agent?

By what it does across a session rather than by what it claims to be. Real agent traffic in this sample read pages and compared products, with about one in seven going on to a successful account action. Abuse touched one route and failed every time.

Should I block unsigned agent traffic now?

Not as a default, in late 2026. Verification is widely deployed but signing is not, so a blanket requirement denies most legitimate automation. Use signatures to grant privileges, and keep behavioural controls deciding the unsigned majority.

Further reading

On this wiki:

Sources: