Urbyte has released test results for its KƎƎLO system covering three AI security scenarios based on incidents involving OpenAI, Anthropic and Hugging Face. The results, it said, show autonomous agents can be blocked from acting beyond authorised limits.
The Australian company ran three controlled scenarios using Google Cloud Firestore persistence and a controlled HTTP provider. Across the tests, it recorded seven unauthorised execution attempts denied before provider invocation, three authorised actions completed with durable receipts, and no additional governed executions from duplicate delivery.
The scenarios were modelled on publicly reported incidents. They tested whether an AI agent that could request an action could also execute it, and whether the system could preserve records of approved or blocked actions across retries and changes in authority.
In the scenario based on the reported OpenAI Medicare statistics portal incident, an agent authorised for public research attempted to access a restricted resource and send an action to an unapproved external destination. Both requests were denied before the provider was invoked, while an authorised research action proceeded and generated a receipt that was reused when the same action was delivered again.
A second test drew on the Anthropic case involving a police homicide webpage. In Urbyte's reconstruction, an agent authorised for research attempted to submit information through an external form and then requested extra submission permission. Both requests were refused, and the system recorded the decision CAPABILITY_NOT_GRANTED.
The third scenario focused on a reported Hugging Face security incident involving an autonomous agent and compromised credentials. Urbyte said an agent credential tried to cross into another tenant's resource boundary and obtain administrative permission beyond its assigned scope. Both requests were denied, and a later attempt after revocation was rejected with the recorded decision GRANT_INACTIVE.
Execution boundary
Urbyte said the tests were designed to highlight the distinction between an AI agent requesting an action and being permitted to carry it out. KƎƎLO, it said, creates a separate execution boundary governing resources, destinations, permissions and organisational scope, while producing durable receipts for approved actions.
The company also ran three bypass controls for comparison. When the enforcement gateway was omitted, the isolated provider accepted prohibited requests that KƎƎLO had blocked in the governed tests.
Urbyte describes KƎƎLO as an Executive Operating System intended to connect organisational objectives, governance, automation and controlled execution. The architecture includes components known as Director and Forge, along with persistent state management and an operating cycle of Detect, Verify, Govern, Act, Measure, Reduce and Prevent.
Jarryd Thamer, Founder and Chief Executive Officer of Urbyte, set out the principle behind the system. "An agent's ability to ask for something is not the same as having the authority to do it. That's the problem KƎƎLO addresses. Organisations need to know what their AI systems are allowed to do, what actually happened, what was prevented and what remains unresolved. Most governance focuses on whether an action was approved. KƎƎLO goes further by governing what happens next. I designed the system around a principle that matters to me personally. Capability should never exceed authority, and authority should never exist without accountability," Thamer said.
Beyond security
Urbyte said the same architecture is also being used in enterprise resource verification, with telecommunications as its first commercial focus. In a controlled assessment involving 5,000 enterprise mobile services across three evidence cycles, it said it identified 266,750.6 GB of verified unused allocation and found 3,512 services that met verification requirements, while excluding 1,488 services that could not be sufficiently verified.
It also disclosed a separate data-centre scenario examining 5,000 infrastructure assets over 84 days. That assessment recorded 4,323 verified assets, 1,157 actionable cases and 1,157 controlled executions with matching receipts, with no duplicate effective executions and no unresolved execution obligations.
Company origins
Thamer linked the company's origins to the Northern Rivers floods, when he said he was unable to access the internet despite holding a large amount of banked mobile data. He founded Urbyte to address what he described as a lack of independent verification and accountability around unused telecommunications allocations, and said the development of an AI-based operating system emerged from that work.
Urbyte said its telecommunications allocation-waste work has entered the Australian parliamentary record through Submission No. 48. The company added that it develops independent allocation-waste verification and is not a telecommunications carrier or data reseller.
Thamer described the wider aim for the technology in the context of autonomous systems in business operations. "The future of AI isn't just about what systems are capable of doing. It's about whether organisations can trust, verify and remain accountable for what those systems actually do. That's what KƎƎLO is here to establish," Thamer said.