Kosa8 Labs has printed a 46-page technical report, “Sealed Machines for Autonomous Agents”, benchmarking 11 AI agent infrastructure platforms towards 13 safety necessities it says are essential for deploying autonomous brokers in regulated environments similar to banks, hospitals, defence suppliers and authorities businesses.
The platforms assessed embrace Docker, E2B, Modal, Cloudflare, AWS and Google. According to the report’s personal scorecard, no rival system absolutely met greater than 4 of the 13 necessities. Kosa8 says its personal platform met all 13, although the corporate wrote and performed the analysis itself, a limitation that deserves editorial weight.
What the report measures
The framework is constructed on a intentionally pessimistic beginning assumption: that an attacker has already gained root entry contained in the agent’s execution setting. From that baseline, the report asks what the compromised agent can nonetheless attain, and whether or not an organisation can reconstruct what it did.
The findings centre on a set of controls that the corporate says most platforms go away unaddressed. These embrace enforced two-person approval for high-risk actions at run time, community controls that block locations outdoors an outlined coverage (together with uncooked IPv6 and cloud-metadata endpoints), and a tamper-evident audit log appropriate for offline verification by a 3rd occasion.
On efficiency, kosa8 reported a median checkpoint time of 638 milliseconds and a rewind time of 520 ms for a operating agent machine, together with reminiscence and processes. Eight assault situations have been examined, protecting exfiltration by way of a poisoned GitHub difficulty, a malicious npm bundle searching for SSH keys, and an MCP device “rug pull”. The firm says kosa8 contained or absolutely dealt with all eight.
The report is candid about shortcomings. Sandbox creation takes round 4 seconds, slower than the warm-pool begins obtainable from hosted cloud providers. One restore path has an unexplained gradual mode. The firm doesn’t but maintain a SOC 2 report or an unbiased penetration check, each of which regulated monetary establishments will sometimes require earlier than procurement sign-off.
“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, founder and chief government of kosa8. “Putting every agent in a sealed machine is now common. What regulated organisations are missing is control and evidence on their own hardware.”
Regulatory and market context
The report arrives at a second when monetary regulators are starting to border expectations round agentic AI. The FCA‘s ongoing AI technique and the Bank of England‘s mannequin threat steerage each emphasise explainability and auditability, necessities that map instantly onto the audit-log and rollback options kosa8 is positioning as differentiators. Under DORA, EU monetary establishments face binding necessities on ICT threat administration and incident response that stretch, in precept, to third-party AI tooling.
The aggressive panorama for AI agent infrastructure in regulated sectors is nascent however filling rapidly. Established cloud suppliers are extending their sandbox and secrets-management providers towards agent workloads, whereas a cohort of specialist startups is constructing purpose-built safe execution environments. Kosa8’s declare to differentiation is the mixture 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 firm is about to launch publicly on Product Hunt on 20 October 2026. Independent verification of the benchmark methodology and a third-party penetration check would be the subsequent credibility milestones the market will look for.
AI stage 1 of 5: written by Darlyn Ho; AI helped with tone, construction or wording; edited and signed off by Mark Walker, Editorial Director. What the levels mean
