Governance · 5 min read · Updated 2026-11-15
When the agent at your door won't say who sent it
Arkose Labs has expanded its Agent Trust Manager to classify and police AI agent traffic at every step of the customer journey. The move answers a problem most enterprises have not yet measured — agents that quietly decline to say what they are.
What happened
Arkose Labs has expanded Arkose Agent Trust Manager, a product for classifying and policing AI agent traffic on customer-facing websites (Help Net Security, 5 November 2026). The expansion lets a business map its customer journey step by step — sign-up, login, checkout, money movement — and assign each step an allow, monitor or block policy, enforced by the company's Arkose Titan engine. Sessions are labelled self-disclosing good, non-disclosing good, or malicious, and a new Arkose Agent ID gives a single agent a stable identifier that persists across network changes, new builds and multiple accounts (Help Net Security, 5 November 2026).
The reason the product exists is a number buried in Arkose Labs' own research. The company's State of Agentic AI Trust Q3 report found that personal AI agents identify themselves less than 3% of the time when they visit a website (Arkose Labs, via Help Net Security, 5 November 2026). As Arkose Labs CEO Kevin Gosschalk put it: "Businesses can no longer count on agents to announce who they are" (Help Net Security, 5 November 2026).
That single sentence describes a shift in the ground rules of web traffic. For twenty years, bot detection has worked by looking for agents that lie about being human. The new problem is agents that are honestly not human and still will not say so — and a business has no reliable way to ask.
Why this is not an isolated incident
Arkose Labs' number lands inside a traffic pattern that was already moving in one direction. Imperva's 2025 Bad Bot Report found that automated traffic made up 51% of all web traffic, the first time in a decade that it has overtaken human activity, with bad bots alone accounting for 37% of all internet traffic (Imperva, 2025 Bad Bot Report). Imperva attributes part of the rise to AI lowering the barrier to building bots in the first place — the same AI capability that now lets a legitimate personal assistant browse, click and buy on a person's behalf without ever announcing itself.
The detection industry's new capability list also tells its own story. Arkose Labs did not add one feature; it added seven in a single release — orchestration by journey step, session classification, branded block messages, a natural-language query interface, a persistent agent identifier, a signal-change audit trail, and cross-network intelligence comparison (Help Net Security, 5 November 2026). A vendor does not build that much tooling in one cycle for a problem it expects to be niche. It builds it because every customer asking is describing the same gap: traffic arriving at checkout and login that nobody can classify with confidence.
What it means for a regulated enterprise
The practical exposure sits at exactly the steps Arkose Labs chose to govern first: sign-up, login and money movement. Each of those is also where fraud controls, know-your-customer obligations and payment-fraud liability concentrate. A business that cannot tell whether a session is a disclosed personal assistant, an undisclosed one, or an attacker probing the same page has three different risk profiles collapsed into one undifferentiated stream, and a single blanket policy — allow everything or block everything — is wrong for at least two of the three.
There is a second exposure that is easy to miss: this is not only about agents arriving at your front door. It is a mirror of the identity-governance problem already building inside the enterprise, where AI agents built in-house hold API keys, OAuth tokens and service accounts that no approval process reviewed. An organisation that cannot say which of its own agents is acting, with what authority, is poorly placed to judge whether an unannounced agent knocking on its website is legitimate traffic or something to stop.
Record-keeping regimes are converging on the same demand from both directions. If an agent — yours or someone else's — touched a regulated transaction, the expectation is increasingly that you can show who or what acted, not merely that you had a policy governing the interaction.
What actually addresses it
The mechanism Arkose Labs is building for inbound web traffic — a stable identifier per agent, a classification recorded at the moment of interaction, and a policy enforced at each step rather than at the perimeter as a whole — is the same mechanism that has to exist on the other side of the relationship, inside the enterprise that deploys its own agents.
A per-agent identity that is issued deliberately rather than inherited gives an organisation a named holder for every action an agent takes, instead of a shared credential nobody can attribute. Scoping that identity to specific tools, data and spend, and logging every use against it, turns "we have a policy" into a record a reviewer can actually read. None of this replaces perimeter tools like Arkose's — the two operate on different sides of the same conversation — but an enterprise that has not built the internal version has no basis for judging what a disclosed or undisclosed agent on the outside is entitled to do once it is let in.
What to check on Monday
Ask your fraud, security and platform teams one question each, without buying anything first: which of our customer-journey steps — sign-up, login, checkout, payment — currently has no way to distinguish a disclosed AI agent from an undisclosed one or from a human. Write down the answer for each step. Most organisations will find the answer is "none of them," and that list, not a vendor demo, is the actual starting point for deciding where policy needs to exist before traffic volume makes the decision for you.
How AANCER answers
AANCER governs the side of this problem that sits inside the enterprise: the agents a business builds and runs itself. Every Certified Agent carries an identity issued by the organisation's own per-install certificate authority, not a borrowed credential, so there is always a named holder behind an action rather than a shared token.
That identity is scoped to the tools, data and spend an agent is meant to touch, and every use is written to an append-only, tamper-evident audit ledger — the record a reviewer needs to answer "what did this agent do, and under what authority" without reconstructing it after the fact. Revocation takes effect in under 30 seconds, so a scoped response to one agent does not require shutting down the systems it was talking to. None of this classifies inbound traffic the way Arkose Titan does; it answers the mirror question, on the side of the relationship an enterprise actually controls.
Sources
- https://www.helpnetsecurity.com/2026/11/05/arkose-agent-trust-manager/
- https://www.imperva.com/resources/resource-library/reports/2025-bad-bot-report/
Related guides
Compliance
The EU AI Act Article 12 readiness guide
What record-keeping and human-oversight obligations actually require operationally from August 2026 — and the evidence an auditor will ask you to produce.
9 min read
Read the guide →Risk
The credentials nobody reviews
Your AI agents hold OAuth tokens, API keys and service accounts that went through no approval process. The agent was reviewed. The studio was reviewed. The identity behind them was not.
5 min read
Read the guide →Security
When the agents organised themselves: what the Hugging Face swarm means for accountability
Roughly 700 AI agents divided labour, traded favours and compromised production infrastructure across four regions. The uncomfortable part is not that it happened — it is that the account of what happened had to be reconstructed afterwards, by outside parties.
6 min read
Read the analysis →