Kosa8 Labs has printed a 46-page technical record, “Sealed Machines for Self sufficient Brokers”, benchmarking 11 AI agent infrastructure platforms towards 13 safety necessities it says are important for deploying independent brokers in regulated environments similar to banks, hospitals, defence providers and executive businesses.
The platforms assessed come with Docker, E2B, Modal, Cloudflare, AWS and Google. In keeping with the record’s personal scorecard, no rival gadget totally met greater than 4 of the 13 necessities. Kosa8 says its personal platform met all 13, despite the fact that the corporate wrote and performed the analysis itself, a limitation that merits editorial weight.
What the record measures
The framework is constructed on a intentionally pessimistic beginning assumption: that an attacker has already won root get admission to within the agent’s execution setting. From that baseline, the record asks what the compromised agent can nonetheless succeed in, and whether or not an organisation can reconstruct what it did.
The findings centre on a collection of controls that the corporate says maximum platforms go away unaddressed. Those come with enforced two-person popularity of high-risk movements at run time, community controls that block locations outdoor an outlined coverage (together with uncooked IPv6 and cloud-metadata endpoints), and a tamper-evident audit log appropriate for offline verification via a 3rd get together.
On efficiency, kosa8 reported an average checkpoint time of 638 milliseconds and a rewind time of 520 ms for a working agent device, together with reminiscence and processes. 8 assault eventualities have been examined, masking exfiltration thru a poisoned GitHub factor, a malicious npm bundle looking for SSH keys, and an MCP device “rug pull”. The corporate says kosa8 contained or totally treated all 8.
The record is candid about shortcomings. Sandbox introduction takes round 4 seconds, slower than the warm-pool begins to be had from hosted cloud services and products. One repair trail has an unexplained sluggish mode. The corporate does no longer but grasp a SOC 2 record or an unbiased penetration check, either one of which regulated monetary establishments will normally require ahead of procurement sign-off.
“Each and every week brings a brand new tale of an AI agent deleting a database or leaking a token, and the business’s solution has been higher activates and smarter filters,” stated Dhruvit Talati, founder and leader govt of kosa8. “Hanging each agent in a sealed device is now commonplace. What regulated organisations are lacking is regulate and proof on their very own {hardware}.”
Regulatory and marketplace context
The record arrives at a second when monetary regulators are starting to body expectancies round agentic AI. The FCA‘s ongoing AI technique and the Financial institution of England‘s style menace steering each emphasise explainability and auditability, necessities that map without delay onto the audit-log and rollback options kosa8 is positioning as differentiators. Below DORA, EU monetary establishments face binding necessities on ICT menace leadership and incident reaction that stretch, in theory, to third-party AI tooling.
The aggressive panorama for AI agent infrastructure in regulated sectors is nascent however filling temporarily. Established cloud suppliers are extending their sandbox and secrets-management services and products towards agent workloads, whilst a cohort of specialist startups is development purpose-built safe execution environments. Kosa8’s declare to differentiation is the mix of hardware-level isolation, run-time coverage enforcement and verifiable audit trails on personal or air-gapped infrastructure, a profile that speaks to corporations cautious of sending delicate workloads to a shared public cloud.
The corporate is about to release publicly on Product Hunt on 20 October 2026. Unbiased verification of the benchmark method and a third-party penetration check would be the subsequent credibility milestones the marketplace will search for.
AI point 1 of five: written via Darlyn Ho; AI helped with tone, construction or wording; edited and signed off via Mark Walker, Editorial Director. What the levels mean