Domain 5 of 6: Be trusted
Trust
You know which agents you deal with and they know who you are: identity, permissions, policies for agents and honesty toward the people behind them.
The outcome and why it matters
As agents gain the ability to act, two questions follow. Who is this agent, and who does it act for? And can the agent trust you? Identity, permission and honesty make action safe for both sides. Without them, the only safe policy is to block, and blocking is a failure of the other domains.
An analogy
Led by: Legal, risk and security. See the responsibility matrix.
Capabilities
5 things an organization does in this domain. Standards named here are explained in the Standards Map.
Agent identity
Knowing which agent is making a request and who operates it, with proof, not just a user agent string anyone can copy.
- What good looks like
- Verified agents are recognized through cryptographic signatures or published address ranges. Unverified traffic is handled under a stricter policy.
- Typical standards and techniques
- Web Bot Auth (IETF work using HTTP Message Signatures, RFC 9421)
- Published IP ranges and reverse-DNS verification used by major crawlers
- Verified-bot programs offered by network providers
Delegated permissions
Knowing what an agent is allowed to do for the person it represents, and limiting it to that.
- What good looks like
- Agents obtain narrow, time-limited access with the user’s consent. Scopes are specific, and people can see and revoke what they granted.
- Typical standards and techniques
- OAuth 2.0 and 2.1 with scopes
- OAuth 2.0 Protected Resource Metadata (RFC 9728)
- OpenID Connect
- Dynamic client registration (RFC 7591)
Agent policy
The organization’s written rules for agents: what they may access, what they may do, what data they may use and what you will not do through an agent.
- What good looks like
- A published, versioned policy is approved by legal, risk and the business, and is enforced in access controls, not just stated in terms of use.
- Typical standards and techniques
- A public agent policy page
- ISO/IEC 42001 and the NIST AI Risk Management Framework for governance of your own agents
- Content and AI-use preference signals
Provenance and disclosure
Honesty toward the people behind agents: saying that content or a reply is generated, and marking where information comes from.
- What good looks like
- Automated replies say they are from an agent. Generated media carries provenance data. Sources are cited and dated.
- Typical standards and techniques
- C2PA Content Credentials for media provenance
- Source citation in answers
- Disclosure notices on automated replies
Security and incident response
Defenses against abuse specific to agents, such as prompt injection through your content, scraping, fraud and impersonation, and a plan when something goes wrong.
- What good looks like
- Agent endpoints are tested for abuse. Logs allow every agent action to be traced to an identity. An incident plan names who acts and how people affected are told.
- Typical standards and techniques
- OWASP guidance for LLM and agent applications
- Audit logging with agent identity
- Rate limiting and anomaly detection
- Incident response runbooks
See the Standards Map for who owns each standard and how mature it is. To learn the skills, see the Specialist certification (Engineering (with legal and risk)).
KPIs
Defined precisely, so two organizations count the same way.
- Verified agent share
- Requests from agents whose identity was verified by signature or address range, divided by all requests from agents that identify themselves.
- Policy coverage
- The share of agent-callable actions that have a documented permission scope, a value limit and an accountable owner.
- Impersonation and abuse incidents
- The count of confirmed cases, in the period, of an agent claiming an identity it did not hold or acting outside its permissions, with time to contain each.
What each maturity level looks like here
- 1
Level 0: Closed
No policy. Agents are blocked or allowed by habit, and nobody can say which is which.
- 2
Level 1: Visible
Basic rate limits and bot protection exist. Agent traffic is treated like any other automated traffic.
- 3
Level 2: Described
A written agent policy exists, and legal and the business have approved it.
- 4
Level 3: Conversational
Agent actions require scoped, revocable permission. Every action is logged to an identity.
- 5
Level 4: Callable
Verified identity is checked on requests. Provenance and disclosure are routine, and incidents are rehearsed.
- 6
Level 5: Trusted
Trust runs both ways: you verify agents and agents can verify you. Reviews and audits are routine.
See all domains side by side or take the self-assessment.