# Arkna

> Arkna is an Australian company that tests AI agents against operating rules and reports findings linked to evidence.

ARKNA PTY LTD is based in Sydney, Australia, and was founded in 2025 by Keith Cheng.

## Service

Arkna agrees the workflow, rules and questions to examine, tests in the customer's environment, and reviews available traces and evaluations. Deliverables are a decision brief, supporting evidence and recommended actions. Scope, fee, access and data handling are agreed before work begins. The customer makes the release decision.

Findings apply to the cases tested and records available. Arkna does not certify safety or compliance. The sample assessment is illustrative work, not a customer engagement; the report states its scope and limitations.

## Company and service pages

- [AI agent testing](https://www.arkna.ai/what-we-do)
- [About Arkna](https://www.arkna.ai/about)
- [Mission](https://www.arkna.ai/mission)
- [Contact](https://www.arkna.ai/book-review)
- [Privacy and terms](https://www.arkna.ai/legal)

## Research

- [Three accounts of the Hugging Face incident](https://www.arkna.ai/research/three-accounts-of-the-hugging-face-incident): A comparison of what OpenAI, Hugging Face and METR’s investigators examined, how the outside review checked its coverage, and which conclusions their evidence supports.
- [Four predictions after the Hugging Face incident](https://www.arkna.ai/research/four-predictions-after-the-hugging-face-incident): four forecasts from the July reports, published as a frozen register with qualifying events, check dates, and pass, fail and not-tested conditions stated in advance.
- [The agent’s version of events](https://www.arkna.ai/research/the-agents-version-of-events): In the same intrusion, one route to changing a system was blocked and another succeeded. Hugging Face’s reconstruction shows how to distinguish an attempted action from its effect.
- [Anatomy of an agent escape](https://www.arkna.ai/research/anatomy-of-an-agent-escape): A package store became a message board. A public application became a launch point. The published reports show how evaluation agents reached Hugging Face’s production systems.
- [Defence in the agent economy](https://www.arkna.ai/research/defence-in-the-agent-economy): Hugging Face had to change models during its investigation. A letter on preparing a usable defensive AI option before a security team needs it.
- [An agent can write to its own record](https://www.arkna.ai/research/an-agent-can-write-to-its-own-record): In the Hugging Face investigation, researchers found tool-call records that misrepresented execution. What would a cryptographic seal preserve in that case?
- [The compounding error problem in production AI](https://www.arkna.ai/research/the-compounding-error-problem-in-production-ai): why benchmark reliability and production reliability are different numbers, and why the production figure has to be measured in production.