Kosa8 Labs Unleashes Critical AI Security Benchmark for Regulated Fintech Firms

Kosa8 Labs has published a 46-page technical report titled “Sealed Machines for Autonomous Agents,” which rigorously benchmarks 11 AI agent infrastructure platforms against 13 crucial security requirements. These requirements are deemed essential for deploying autonomous agents in highly regulated environments, including banks, hospitals, defense suppliers, and government agencies. The platforms evaluated in the report include prominent names such as Docker, E2B, Modal, Cloudflare, AWS, and Google.

A key finding from the report’s own scorecard indicates that no rival system managed to fully meet more than four of the 13 stipulated security requirements. In contrast, Kosa8 asserts that its proprietary platform successfully met all 13 requirements, though the company acknowledges that it conducted and wrote the research itself, a factor that warrants editorial consideration.

Dhruvit Talati, founder and CEO at Kosa8, explained that the framework for the report is built on a deliberately pessimistic starting assumption: that an attacker has already gained root access inside the agent’s execution environment. From this baseline, the report investigates what a compromised agent can still access and whether an organization can effectively reconstruct its actions. The findings primarily center on a set of controls that Kosa8 contends most other platforms leave unaddressed.

These critical controls include enforced two-person approval for high-risk actions at runtime, robust network controls that block destinations outside a defined policy (encompassing raw IPv6 and cloud-metadata endpoints), and a tamper-evident audit log suitable for independent, offline verification by a third party. On the performance front, Kosa8 reported a median checkpoint time of 638 milliseconds and a rewind time of 520 ms for a running agent machine, inclusive of memory and processes. The company also tested eight distinct attack scenarios, covering threats such as data exfiltration through a poisoned GitHub issue, a malicious npm package designed to hunt for SSH keys, and an MCP tool “rug pull,” claiming that its platform contained or fully handled all eight.

The report is candid about certain shortcomings of the Kosa8 platform. Sandbox creation, for instance, takes approximately four seconds, which is slower than the warm-pool starts offered by hosted cloud services. Additionally, one specific restore path exhibits an unexplained slow mode. Kosa8 also acknowledges that it does not yet possess a SOC 2 report or an independent penetration test, both of which are typically mandatory for procurement sign-off by regulated financial institutions.

“Every week brings a new story of an AI agent deleting a database or leaking a token, and the industry’s answer has been better prompts and smarter filters,” stated Dhruvit Talati. “Putting every agent in a sealed machine is now common. What regulated organizations are missing is control and evidence on their own hardware.”

This report arrives at a crucial juncture as financial regulators begin to formalize expectations around agentic AI. The Financial Conduct Authority’s (FCA) ongoing AI strategy and the Bank of England’s model risk guidance both underscore the importance of explainability and auditability—requirements that directly align with the audit-log and rollback features Kosa8 is positioning as key differentiators. Furthermore, under the Digital Operational Resilience Act (DORA), EU financial institutions face binding requirements concerning ICT risk management and incident response, which extend to third-party AI tooling.

The competitive landscape for AI agent infrastructure within regulated sectors is currently nascent but rapidly expanding. Established cloud providers are actively extending their sandbox and secrets-management services to accommodate agent workloads, while a growing cohort of specialist startups is developing purpose-built secure execution environments. Kosa8’s unique claim to differentiation rests on its combination of hardware-level isolation, real-time policy enforcement, and verifiable audit trails, all deployable on private or air-gapped infrastructure. This profile specifically targets firms hesitant to send sensitive workloads to a shared public cloud.

Kosa8 is scheduled to launch publicly on Product Hunt on October 20, 2026. The next critical credibility milestones that the market will anticipate include independent verification of its benchmark methodology and a third-party penetration test.