NEAR AI × Vijil: certifiable trust for agentic transactions

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 cryptocurrency.

The problem

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 to 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 four proofs.

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.

Verifiable trustworthiness at scale

NEAR AI established the mandate that 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.

Vijil as a trust certification engine

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. As the NEAR AI Auditor Agent, Vijil Diamond operates as a Certificate Authority, issuing and revoking on-chain certificates assigned to certified 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 is running, not a spoof;
• the trace and results, as part of the attestation evidence provided from the TEE, give mathematical backup for certification;
• the agent code and 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, and 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, with certificate chaining and the Auditor Agent serving as the certificate authority.

3. Regulatory compliance by design

Enabled demonstration of compliance with multiple regulatory frameworks using a shared testing infrastructure, and 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.

Read this case study as a PDF →

Diamond runs the same way on your own agents. Start for free, ortalk to us about an evaluation against your policy.