Skip to the yard
Waxsealai

AI Agent Access Needs Identity, Reputation and Accountability

Shipped
Shipper
Bryan Becker
Readers
392

Cloud hosted agents can scale web activity while obscuring who they act for. New standards and site tools aim to restore visibility and control.

AI Agent Access Needs Identity, Reputation and Accountability

Cloud hosted AI agents can make web services harder to serve safely: they can act at a person’s direction, but a site may see only the agent and its operator. As these systems take on more browsing and transactions, sites need ways to understand which agent is involved, assess its activity and connect consequential actions to the person responsible. That visibility could let sites accommodate more agent traffic while managing abuse, fraud and capacity.

Personal agents take a user’s instructions and pursue a goal by controlling a browser or cloud computer. They are often hosted and maintained by an operator as part of a managed service or consumer product. Examples named in this discussion include Muse from Meta and dots from OpenAI. Since agents can run continuously, act in parallel and tolerate repeated friction, they can make requests at a scale that differs from ordinary browsing. Many websites’ capacity planning and business models were built around the limits of human visitors.

When the agent becomes the only visible visitor

A remotely hosted agent acts as an intermediary: it sits between a user and a site, so the site receives requests from the intermediary rather than directly from the person. Cloud browsers and traffic relays such as proxies and VPNs have created similar gaps in visibility. The new concern is the potential scale of agent activity. If blocking intermediaries becomes impractical as agent traffic grows, sites must decide whether and how to trust them without the direct context they once had about visitors.

That gap affects ordinary service as well as security. A site may have trouble distinguishing the preferences of several people whose requests arrive through one agent. It may also struggle to determine who authorized a disputed purchase, or to stop suspicious activity tied to one person without disrupting every other person using the same service.

Consider Ezra, a personal assistant who runs errands for four clients. A grocer once recognized Amy, one of those clients, and could tailor service to her. With Ezra handling the orders, the grocer sees one visitor and has less insight into which client wants what. If Bob later disputes a phone purchase Ezra made, the store may have no record connecting Bob to the request. If Charlie asks Ezra to try a list of passwords at his bank, the bank may see suspicious attempts but be able to identify only Ezra. Blocking her would also affect Amy, Bob and Debra.

The analogy captures the core issue for websites: an intermediary can bring useful business while hiding the person behind a request. Sites bear much of the cost when they cannot distinguish legitimate actions from harmful ones, resolve disputes or apply limits to the right account. An agent can also repeat an action at a scale no individual could manage, increasing the impact of behavior that might otherwise be small.

The discussion focuses on agents hosted away from the user’s device. Locally run agents, including OpenClaw, and agents operating in a person’s own browser raise different questions. In those cases, there may be no single operator to identify or hold responsible, and this work does not make them its initial focus.

Preserve the relationship between site and user

The guiding principle is that intermediation should not erase the business relationship between a user and a site or shift each participant’s security responsibilities. Sites need a way to recognize and serve people acting through agents, and to hold the relevant party accountable when needed. They also need to recognize the agent and its operator, since automated systems can act at a scale beyond an individual user.

This goal does not require agents to imitate people or a person to approve every step. A site should be able to request information appropriate to an action, whether the person is browsing directly or using an agent. The information might be minimal for routine access and more specific for a sensitive transaction. A person should still be able to browse privately, while a bank may require additional verification before allowing account access.

Identity, reputation and accountability answer different questions. Identity helps establish who or what is making a request. Reputation helps a site judge whether that entity’s activity is trustworthy. Accountability identifies who answers when an action causes harm. A useful system needs to support all three without demanding the same disclosure for every request.

Why common workarounds leave gaps

One proposed fix is to pass a CAPTCHA to the user. That shows a person was present at one moment, but it does little to distinguish one client from another or establish who stands behind the agent’s later actions. Human approval can help prevent an agent from acting against its user’s intent, but it does not solve the broader visibility problem between a site and the people served by the agent. It also reintroduces friction that agent operators may have incentives to reduce.

Access controls that limit an agent’s permissions have a related shortcoming. Agents are intended to carry out a wide range of tasks, so users may grant broad access. A permission covering everything gives the site little useful information. Even a narrow permission describes what the agent may do; it does not say who is responsible if the agent makes a mistake. Sites need both rules for permitted actions and a way to assess the request and its accountability.

Protocols and identity at different levels

Sites and agents need shared ways to negotiate which actions are permitted and to communicate site preferences. Those tools must be practical for a range of sites and agent services to support. Existing standards may cover some needs, while other gaps may call for new protocols.

One effort, MOLE, aims to support rate limits without relying on user specific identifiers. That could help sites move from implicit recognition toward explicit protocols. The x402 protocol similarly aims to support web payments while preserving privacy. These approaches point toward ways to set limits or complete transactions without exposing a person’s identity by default.

For identifying an agent and its operator, the preferred approach described here is Web Bot Auth (WBA). It lets an agent sign requests, giving a site a way to verify the agent’s identity and, implicitly, the operator behind it. A remaining problem is nested intermediation. If Agent A uses Agent B inside Browser C, a site needs to understand that chain so it can apply access rules correctly and detect attempts to get around them.

Agent identity does not necessarily identify the user. Depending on the task, a site could need a session identifier, an account link or a user specific identifier. The right level depends on what the site is trying to decide, and minimizing personal information remains a design goal. The options form a spectrum:

  1. No identifier: A user can make requests without presenting an identifier. Privacy preserving tools such as Privacy Pass and MOLE are intended to help sites assess whether access is acceptable without learning who the user is. Private Access Tokens were an earlier attempt at establishing trust without identifying a user; MOLE is intended as a successor.
  2. Session: A site can link requests within one session to a single agent and build a reputation for that session. The Poppy protocol, for example, includes a session ID for this purpose.
  3. Account: A site can link activity across sessions to the same user, potentially through a stable identifier provided by an operator or another service. This could support linking a site account to an agent account.
  4. Unique identifier: For higher risk actions, scarce resources or stricter verification needs, a site may require an identifier that persists across operators, accounts, agents and sessions. Identity verification services and decentralized proof of personhood are examples of approaches that could provide this type of signal.

The site ultimately decides what information it needs to determine whether an entity may take a particular action at a particular time. Different standards can carry different pieces of that information: WBA for an agent’s platform identity, MOLE for user reputation and OAuth for account identity. The challenge is to make these signals work together without turning every routine request into an identity check.

Claims and site instructions

Some actions depend on a fact about a person rather than a persistent identity. A site might need to know whether someone is over 18 or is the cardholder. An operator may already have checked that fact during sign up, or a third party may hold the relevant information. Sites need a way to request a claim and for operators or third parties to provide it, without requiring agents to expose more information than the action calls for.

Sites also need a machine readable way to explain what agents may do. Today, an agent often discovers restrictions by trying an action and interpreting the response, such as a block. That trial and error can waste resources and make well behaved agents look suspicious. Robots.txt provides instructions for crawlers, but does not fully describe the range of actions a personal agent might take.

The Internet Engineering Task Force’s AI Preferences Working Group is working on ways to express preferences, with an emphasis on uses such as training, inference and search. The broader need is to describe agent access and actions across a site. That work does not yet have a dedicated group focused on the full set of preference mechanisms proposed here.

Open questions for local agents

Agents running on a user’s device or on infrastructure the user controls present a different accountability problem. They look like automation, but may lack a single operator that can build a reputation or answer for their behavior. One possible response is to require a unique identifier, but addressing local agents is outside the initial focus of Trusted Agent Delegation.

From standards work to deployed controls

Interoperable standards will matter because site specific mechanisms can fragment the experience for both agents and websites. The effort described here involves operators, identity verification services, payment providers, merchants, publishers, proxy providers and cloud browsing companies. Its stated aim is to connect operating experience with work on WBA, AI preferences, OAuth, WIMSE and MOLE, including by refining open protocols and deploying implementations.

The Trusted Agent Delegation (TAD) coalition is being formed to bring participants together, examine existing protocols and clarify which problems may need further standards work. Planned activities include meetings alongside IETF events and workshops. The effort is aimed at people building agents and runtimes, operating edge or cloud platforms, and managing commercial or compliance systems.

Some controls are already available to Cloudflare customers. Security Analytics now identifies Personal Assistants as a traffic category and includes a widget showing which agents visit a site, their request counts, paths and errors. Enterprise Bot Management customers can use Business Insights to examine AI crawler and bot activity, including how much each operator crawls compared with the traffic it sends back.

Through AI Crawler and Agent access policies, a site can set access by agent: allow it site wide, restrict it on pages with ads or block it across the site. Search and training preferences remain separate from these controls.

For sites using Account Abuse Protection in early access, login and sign up activity can show whether a personal agent made the request and identify the agent. If one account is abused through a shared agent, a site can take action on that account using its Hashed User ID in a WAF rule, without blocking the agent for every other user. That offers a practical response to traffic commingling: separate the account involved from other people who rely on the same intermediary.

These tools give sites some immediate visibility while standards continue to develop. The larger test is whether shared protocols can let sites state their requirements, agents communicate who they represent and operators take responsibility in ways that work across services. That combination would give sites more control over risk while preserving access for legitimate agent use.

Source: Bryan Becker · blog.cloudflare.com

Same vesselRelated cargo