Beyond the Black Box: Why Auditing AI Agents Requires a New Blueprint for Trust

CHARLESTOWN, MA – September 22, 2026 – As artificial intelligence permeates the administrative arteries of modern society—from adjudicating healthcare claims to managing our social safety nets—a profound crisis of trust has emerged. We are handing non-deterministic systems the keys to highly regulated, deterministic environments, and our traditional methods of corporate oversight are failing to keep pace. Today, a Charlestown-based infrastructure startup unveiled a technical framework that exposes exactly why legacy compliance is broken, and more importantly, how the industry can fix it.

Multikor.ai, Inc. has introduced an agentic enterprise data fabric designed specifically for regulated industries. But underneath the dense technical nomenclature lies a fundamental philosophical shift in how we govern autonomous systems. The company is moving the industry away from static, point-in-time compliance documents and toward continuous, cryptographically secure evidence generated in real time as the system operates.

For years, the intersection of technology, public policy, and corporate responsibility has been haunted by the "black box" problem. When an AI agent denies a medical prior authorization, flags an insurance claim, or routes a patient's protected health information, understanding the "why" is not just a matter of customer service; it is a legal and ethical imperative. Multikor’s architecture suggests that we have been trying to solve this modern problem with obsolete tools.

The Determinism Fallacy and the Death of the Static Audit

Every major cybersecurity and authorization framework currently in use—including the National Institute of Standards and Technology's Risk Management Framework, Software Bills of Materials (SBOMs), and triennial Authority to Operate (ATO) accreditations—relies on a core assumption: software is deterministic. If you inspect the code, scan its dependencies, and verify its configuration, you can guarantee what it will do in production.

An autonomous AI agent shatters this assumption entirely.

A static SBOM reveals absolutely nothing about an agent's dynamic decision pathways. A model card describes the training data and performance benchmarks, not the specific actions an agent takes when querying a live patient database or invoking an external application programming interface. Recognizing this massive vulnerability, federal agencies and the Department of Defense have begun shifting toward Continuous Authorization to Operate (cATO), demanding real-time behavioral telemetry. Yet, the enterprise software market has largely responded by slapping external guardrails onto existing foundation models and calling it governance.

"Most of what is marketed as guardrails is guidance," said Suresh Nelakantam, Co-Founder and Chief Executive Officer of Multikor. "A guardrail is what the platform enforces when it does not. Health systems and insurers are not asking us for smarter agents. They are asking for enforced limits and refusals on the record, because that is what their compliance teams will sign."

The startup addresses this by shifting governance directly to the execution runtime. Its data fabric resolves an agent's boundaries at load time, mapped against the specific topology of the tenant. If an agent attempts to step outside its permitted surface, the system does not simply flag the error for a later review; it refuses to run. There is no degraded mode and no override flag available to developers. Every action—what was requested, what was permitted, what was refused, and under whose authority—is written to an immutable ledger as a native property of execution.

The 24-Hour Blind Spot in Cloud Economics

Beyond the compliance breakdown, the deployment of agentic AI has exposed a massive vulnerability in enterprise cloud economics. While major hyperscalers have made strides in cost attribution over the past year, real-time control remains virtually nonexistent.

According to infrastructure analysts monitoring cloud billing latencies, platforms like Amazon Web Services and Microsoft Azure can take up to 24 hours to reflect token consumption in their enterprise billing dashboards. In an era of autonomous agents, that latency is financially dangerous. If an agent enters a recursive loop while attempting to parse a complex medical record or reconcile an insurance policy, it can burn through thousands of dollars in token fees before a human FinOps manager ever receives an automated alert.

"Attribution was the easy half, and it arrived," said Leigh Turner, Co-Founder and Chief Technology Officer at Multikor. "You can now learn precisely who spent what, a day later, with no mechanism that could have stopped it. That is a receipt, not a control. So we stopped instrumenting the token and moved the decision to the only place it can be refused."

To bypass this hyperscaler blind spot, the company developed a Small Language Model (SLM)-first architecture utilizing a proprietary mixture-of-experts design. Instead of routing every query to expensive, generalized frontier models hosted in the public cloud, the system processes the vast majority of tasks locally. Running on an owned four-node NVIDIA GB10 fleet with paired 200-gigabit interconnects, this approach drastically alters the economic reality of AI deployment. The company reports its monthly infrastructure spend for production agents across multiple tenants is under $2,700—a stark contrast to the $20,000 to $50,000 typically reported for comparable workloads running on hyperscaler infrastructure.

Engineering for Zero Trust and Human Oversight

In sectors like healthcare and insurance, the stakes of automation are deeply human. A misrouted data request or an unverified tool call can result in severe privacy violations, denied care, or compromised social safety nets. To build systems that actually allow communities to thrive rather than merely extracting efficiency, the underlying infrastructure must be designed for zero trust.

One of the most complex challenges in this space is reconciling the right to privacy with the demand for rigorous auditability. Health systems are bound by regulations like HIPAA to redact or destroy protected health information upon patient request, yet compliance frameworks require immutable proof that the data was handled correctly before it was destroyed.

The newly detailed architecture solves this paradox through a dual erasure-and-audit mechanism. Redacted values are isolated in a tenant-partitioned store addressable by field. When a value is destroyed, the sensitive plaintext vanishes, but the ledger retains cryptographic proof that the data existed, its classification category, and the exact record of its redaction. Every reveal of that data writes its own permanent audit row.

Furthermore, the system binds human oversight directly to the tool call. If an agent requires human approval to proceed with a sensitive action, that approval is cryptographically tied to the canonically sorted arguments of the request. It cannot be manipulated or widened by re-proposing the action. Circuit-breaker counters restore across the hold, and a stop recorded during a pause cannot be overridden by a later approval from a different administrator.

"The measure of an architecture is not how it performs on the day you ship it," said Kimberly Boydston, Co-Founder and Chief Architect. "It is how much of it you have to touch when the requirement changes. A standard that only exists in a document is a suggestion with better formatting—so we put ours somewhere good intentions cannot reach it."

A Substrate for Systemic Accountability

As we integrate artificial intelligence deeper into the systems that govern our collective future, we can no longer afford to treat governance as an afterthought or a bolt-on accessory. The transition from static documentation to continuous, cryptographically enforced evidence is not just a technical upgrade; it is a necessary evolution in corporate responsibility.

By anchoring authorization to the data fabric itself, and by utilizing advanced high-radix switching to ensure low-latency enforcement without sacrificing oversight, this new framework proves that accountability can be engineered directly into the substrate of our technology. The platform also matures its trust dynamically—warning during bootstrap phases, soft-rejecting during calibration, and strictly hard-rejecting in full production as evidence of safety accumulates.

For the enterprise risk officers, federal compliance auditors, and everyday patients relying on these systems, the true promise of artificial intelligence can only be realized when its boundaries are as intelligent, resilient, and unyielding as the models themselves. The era of the unauditable black box is closing, making way for an infrastructure where evidence is not just requested, but fundamentally guaranteed.