NEAR AI provides a highly-secure environment for the execution of workloads and inference using sensitive data, supporting automated multi-agent transactions on behalf of users such as the sale and transfer of crypto currency.

Before being deployed on the NEAR AI platform, security and engineering teams need systematic verification of stringent trustworthiness proofs for third-party agents. Gaps in the evaluation and verification processes could expose the platform to agents executing rogue transactions, agents compromising systemic integrity through code that only executes at runtime, or violating data privacy commitments at inference time - with negative business and financial repercussions.

To offset the risks involved with accepting third-party agents to run in its environment and minimize the potential for financial loss through the actions of rogue agents, NEAR AI required a systematic and repeatable process for evaluating agents for approval prior to deployment, encompassing:

‍Proof of Trust

‍Business, security, and compliance stakeholders required systematic verification to consistently ensure agents were neither malicious nor vulnerable to exploits.

‍Proof of Confidentiality

‍Before deploying the agent, stakeholders needed a mechanism to verify that the code tested in the trusted environment was the same as the code being deployed — and that the testing process and system itself was uncompromised.

‍Proof of Resilience

‍To ensure defense in depth, security teams needed validation of platform guardrails to verify that they would detect attacks at runtime.

‍Proof of Regulatory Compliance

‍Compliance teams required proof that agent actions were compliant with regulations such as the EU AI Act.

Traditional approaches to testing agents failed to take into account that agents (especially those with elevated permissions) can reason and act independently in ways that are non-deterministic. Shallow or static testing without any integrity or confidentiality guarantees of the code's chain of trust proved insufficient for trustworthiness assurance, and to maintain the velocity required by NEAR AI to meet its business goals.

Compliance became something teams argued about, not something the system could prove.

NEAR AI established the mandate that every agent must be uniformly trustworthy, resilient, compliant and verified to run on its platform.

Every agent must be uniformly trustworthy, resilient, compliant and verified to run on its platform.

This required a certification gate performed in a trusted execution environment (TEE) that could be applied consistently across all agents, teams, and development environments to securely evaluate agents before deployment.

NEAR AI partnered with Vijil to deliver an Auditor Agent to automatically certify submitted agents in a TEE, using a set of bespoke test harnesses. The Diamond agent provided behavioral, static, and dynamic testing. The output from the evaluation process produced a report that stakeholders could assess to determine trustworthiness. Using Vijil, the NEAR AI platform could provide systematic proof output of agent trustworthiness to internal and external human stakeholders, as well as agent stakeholders. As the NEAR AI Auditor Agent, Vijil Diamond operates as a Certificate Authority, issuing and revoking on-chain certificates assigned to trusted agents.

The report, and the process itself, running in a confidential computing environment, could be embedded in the attestation evidence the environment provides to prove:

• the environment is secure and verifiable
• the real Vijil auditing code (agent) itself is running (not a spoof)
• the trace and results being a part of the attestation evidence provided from the TEE provide mathematical backup for certification
• the agent code or data itself is never compromised
  ‍

1. Agent Trustworthiness Testing

‍Custom test harnesses to certify the behavior of the NEAR AI Auditor Agent. Evaluations included:

• Behavioral consistency across various scenarios
• Reliability under adversarial conditions
• Compliance with declared intent
• Dynamic analysis of behavior for potential malicious functions

2. Proof of Integrity

‍The Auditor Agent performed evaluation in a trusted execution environment, allowing security and engineering teams to verify the chain of proof:

• Evidence that the code tested in the trusted execution environment is the same code that is deployed
• Certificate chaining, with the Audit Agent serving as the certificate authority

3. Regulatory Compliance by Design

‍Enabled demonstration of compliance with multiple regulatory frameworks using a shared testing infrastructure, minimizing blind spots for security and legal teams.

• Automated production of a report for audit and review by compliance stakeholders.

For security and legal teams, this meant fewer blind spots and less regulatory fragmentation.