Urbyte Tests KƎƎLO Security System in Three Scenarios
Urbyte released test results showing its KƎƎLO system denied unauthorised AI agent actions before any provider was invoked.
Updated on Nov. 13, 2026 in Artificial Intelligence

Urbyte released test results for its KƎƎLO AI security system covering three scenarios, recording seven unauthorised execution attempts denied before provider invocation and three authorised actions completed with durable receipts. The scenarios were modelled on publicly reported incidents involving OpenAI, Anthropic and Hugging Face.
Why it matters
The results show how an enforcement layer can refuse AI agent requests that exceed granted permissions, including access to restricted resources and unapproved external destinations. The tests were controlled runs, so their relevance to live deployments is not yet established.
In the controlled tests, KƎƎLO ran on Google Cloud Firestore persistence with a controlled HTTP provider, and the company ran three bypass controls in which the enforcement gateway was omitted. When the gateway was omitted, the isolated provider accepted prohibited requests that KƎƎLO had blocked in the governed tests.
The players
Urbyte
An Australian company founded in response to the Northern Rivers floods, which developed the KƎƎLO AI security system.
The details
In one scenario, an agent authorised for public research attempted to access a restricted resource and send an action to an unapproved external destination, and both requests were denied before the provider was invoked. In another, an agent authorised for research attempted to submit information through an external form and requested extra submission permission, with the system recording the decision CAPABILITY_NOT_GRANTED.
Timeline
10 November 2026: Urbyte released the KƎƎLO test results.
84 days: Duration of the data-centre assessment covering 5,000 infrastructure assets.
Roadmap
Urbyte's KƎƎLO places an execution boundary between AI agents and external resources, rather than relying on the models' own controls. The scenarios tested were modelled on publicly reported incidents involving OpenAI, Anthropic and Hugging Face.
For users of AI agents, a boundary that denies out-of-scope actions could limit unintended data sharing or access to restricted resources. The tests were controlled, so the practical protection in everyday use has not yet been demonstrated.
The takeaway
Tools that enforce permissions outside the AI model itself may be a key part of how agent deployments are secured. Readers should watch for independent evaluation beyond controlled test scenarios.
Further reading
Explore more Artificial Intelligence coverage to see how agent security tools are being developed.
Source note: This article includes information reported by IT Brief Australia.






