The Identity Crisis At The Heart Of The Agent Economy
Rohan Pinto is CTO/Founder of 1Kosmos and a strong technologist with a strategic vision to lead technology-based growth initiatives.
gettyEvery time an AI agent introduces itself with a name, model version or friendly handle, it makes a claim that proves little on its own. Anyone can copy a name or impersonate another agent. In a world where autonomous agents negotiate contracts, move money, delegate tasks and act on behalf of people and organizations, a self-declared identity is not enough. The claim must be verifiable, and the authority behind it must be enforceable.
The agent economy faces a problem familiar from the early internet. Caller ID appeared trustworthy until technology made it easy to spoof a number. Agent networks face a similar challenge. The question is no longer whether an agent can claim an identity. It is whether that identity can be verified and trusted despite stolen credentials, compromised runtimes, malicious delegation and model changes.
Cryptography provides the foundation for trustworthy agent identity. A signature over a fresh challenge can demonstrate control of a private key. A trusted enrollment process can bind that key to a registered agent and its accountable sponsor.
But control of a key does not establish good intent, prove that the execution environment is uncompromised or authorize a specific action. Identity and authorization are separate layers. Identity helps establish who is making a request. Authorization determines whether that actor may perform a particular operation against a specific resource.
An agent can have a valid identity and still attempt an unauthorized transaction. It can also be manipulated through prompt injection or compromised tools into requesting an action its sponsor never intended.
Identity establishes a basis for trust. It does not eliminate the need to enforce boundaries.
Trustworthy agent identity requires resistance to credential theft and cloning, tamper evident records, protection against impersonation and verifiable accountability.
These are security objectives, not absolute guarantees. Public keys and identifiers can be copied without compromising the associated private key. The real concern is whether an unauthorized party can use that information to impersonate the agent or obtain its authority.
Hardware-backed keys, secure enclaves, trusted platform modules and managed key services can make private key extraction substantially harder. Attestation can provide evidence about the execution environment, while short-lived credentials can limit the exposure window after a token is compromised. However, a compromised runtime may still abuse a legitimate signing interface.
The goal is not to promise an identity that can never be compromised. It is to make impersonation harder, constrain misuse and ensure that security failures can be detected and attributed.
Consider Agent A delegating a task to Agent B, which then calls Agent C to access sensitive information. If all three use the same credentials, investigators may struggle to establish which agent initiated an operation or whether the delegation was authorized.
Real delegation must preserve the relationship between the original authority and the downstream action. Each grant should define the permitted scope, recipient, duration and relevant conditions. An agent should receive only the authority necessary for its task, not unrestricted access inherited from another agent.
Recording a delegation chain is not enough. The relevant enforcement point must validate the authority before permitting access.
Revocation matters just as much. It should prevent new authorization and limit the usefulness of credentials already issued. Depending on the architecture, it may also stop work in progress. It cannot undo a completed transaction.
Short-lived credentials, runtime authorization checks, cancellation mechanisms and tamper evident audit records help reduce exposure. A log supports accountability, but only enforcement mechanisms can prevent unauthorized operations.
AI models, context windows, tools and execution environments can change frequently. If identity is tied directly to model weights, every model update risks disrupting continuity and historical accountability.
A stronger approach attaches logical identity to the registered agent and its accountable sponsor, rather than to a particular model checkpoint.
However, continuity of identity must not mean continuity of trust without reassessment. The system should preserve the agent’s identity while recording the model version, runtime, relevant configuration and policy revision associated with consequential operations. Material changes may require additional evaluation or approval.
The identity of the accountable actor and the provenance of the system performing the work are related, but they are not interchangeable.
If an AI agent makes a million-dollar mistake, accountability should not begin with a forensic investigation after the damage occurs.
Before receiving consequential permissions, an agent should have an established sponsor and approval path. For important operations, the evidence chain should connect the agent identity, execution context, delegation, authorization decision, resource accessed and resulting action.
Cryptographically protected records can help establish what the system permitted and which registered identities participated. They do not automatically determine legal liability, which depends on applicable law, contracts and circumstances. They provide a stronger evidentiary foundation for accountability.
A named sponsor without enforceable boundaries is insufficient. A secure key without attributable operations is insufficient. An audit trail without reliable evidence is insufficient. Trustworthy systems need all three: accountable sponsorship, enforceable limits and verifiable evidence.
Organizations evaluating agent identity platforms should examine how credentials are protected, sponsorship is established, delegation is constrained, authorization is enforced, revocation is handled and evidence is preserved across model changes. They should distinguish published design principles from implemented capabilities, evaluating security claims against actual deployment configurations and available evidence.
The agent economy requires more than names or keys. It requires verifiable identity, enforceable authorization and delegation that preserves accountability. An identity that can be impersonated without detection is not trustworthy. An identity without enforceable boundaries is not enough to make an autonomous agent safe.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
/https://i.s3.glbimg.com/v1/AUTH_37554604729d4b2f9f3eb9ad8a691345/internal_photos/bs/2026/Z/E/37imQ5RyiBh5JU7NiyBQ/foto11emp-101-medias-b4.jpg)
/https://i.s3.glbimg.com/v1/AUTH_37554604729d4b2f9f3eb9ad8a691345/internal_photos/bs/2022/G/2/dpESlOTpWU4r9AbvuRmA/22esp-100-indus-a18-img02.jpg)
