AI Inquiry Atlas

An unofficial, independent guide to the public records of the Joint Select Committee on Artificial Intelligence. Not connected with the Parliament of Australia or the committee. Its makers also made a submission (no. 481; see About). Summaries and scores are written by AI and can be wrong: check the original before relying on them.

Submission 409

Orvelis

Submitter group: Companies.

An AI model identified the submitter group and wrote the summary, key points and recommendation summaries. Counts do not measure importance or merit.

Open in the interactive guide

Summary

Read the original submission for the full text.

Orvelis, a company that develops AI harness technology and says it would be among the parties affected, argues the 'AI harness', the control layer above a model, is what organisations and regulators can most directly govern. It makes eight recommendations: a harness standard, government sandbox, procurement parameters, AI Safety Institute assurance, telemetry and incident reporting, sovereign capability and assurance technology.

Key points

Quotes are the submitter’s own words. Text taken from documents, including scanned pages read by text-recognition software, can differ from the PDF. Check the original. Listed in the same order as on the submission’s page in the interactive guide.

  1. Describes the AI harness as the layer that combines a model's judgement with deterministic tools, safety controls and a person on the loop.

    That component is the AI harness: the layer that combines a model's judgement with deterministic tools, safety controls and a person on the loop.

  2. Argues that organisational policy for what agents may do is written in the harness rather than the model, making it a practical point of control.

    The harness, not the model, is the surface on which an organisation writes its policy for what its agents may and may not do.

  3. Argues that Robodebt's lessons on oversight, transparency, record-keeping and accountability apply to agentic AI, and that these are things a harness governs.

    The lesson generalises to agentic AI because the Commission's findings concerned oversight, transparency, record-keeping and accountability rather than any particular algorithm, and those are precisely the things a harness governs.

    A program corrected obvious text-recognition errors in this quote for display.

  4. Argues that it is the deployed system, rather than the model in isolation, that produces the risk.

    it is the deployed system rather than the model in isolation that produces the risk.

  5. Argues that because the harness outlasts individual models, it is where durable sovereign capability can be built cost-effectively.

    Since the harness is the enduring layer while models are replaced, this is where durable sovereign capability can be built on a cost-effective basis.

  6. Argues that assurance technology could be a field of national capability, as assuring deployment does not require owning the largest model.

    A country does not need to own the largest model to assure how models are deployed, and assurance is a market in its own right. Positioning AI assurance technology as a deliberate field of national capability would give Australian suppliers an export edge

Recommendations

Recommendations are shown beside the submission’s own words. Listed in the same order as on its page in the interactive guide.

  1. Commission Standards Australia to define AI harness building blocks and minimum capabilities: context, tools, memory, policy loading, session/state control, safety/permission gates, subprocess coordination, and audit/observability.

    That the Government commission Standards Australia to develop an Australian Standard for AI harnesses that defines their constituent building blocks and the minimum capabilities each must provide. The Standard should address the components common to agentic AI systems, set out in Table 1: assembly of the context supplied to the model, access to tools and systems, memory, the loading of standing instructions and organisational policy, session and state control, permission and safety gates, coordination of sub-processes, and audit and observability.

  2. The Standard should require interoperability and component portability, so a harness is an open integration fabric rather than a closed product tied to a single vendor.

    It should require interoperability and component portability, so that a harness can be built as an open integration fabric into which tools and utilities from many developers are composed, and from which any one component can be replaced, rather than as a closed product tied to a single vendor.

  3. Have the Standard accommodate both software-based and hardware-based enforcement, so it applies to embodied or intermittently connected systems in critical infrastructure and defence as well as enterprise deployments.

    It should accommodate both software-based and hardware-based enforcement, so that it applies to embodied or intermittently connected systems used in critical infrastructure and defence as well as to conventional enterprise deployments.

  4. Have the Australian Standard profile and adapt existing international work such as the OWASP Agent Control Standard and NIST control overlays rather than begin from a blank page.

    Where international work already covers this layer, such as the OWASP Agent Control Standard24 and the control overlays NIST is developing for AI agent systems25, the Australian Standard should profile and adapt that work rather than begin from a blank page.

  5. Require in the harness standard that the assurance functions of a harness be inspectable by a party other than its supplier.

    A standard for harnesses is the natural place to require that the assurance functions of a harness be inspectable by a party other than its supplier.

  6. Have the Digital Transformation Agency operate a standing sandbox with representative Commonwealth datasets (de-identified or synthetic where required) for industry and researchers to develop and test AI harness capabilities for public-sector use.

    That the Digital Transformation Agency establish and operate a standing sandbox environment, provisioned with representative Commonwealth agency and departmental datasets, de-identified or synthetic where required, in which industry and researchers can design, develop, test and optimise AI harness capabilities for public-sector use under controlled conditions.

  7. Require all Commonwealth AI tender requirements to specify mission-specific harness parameters and deliverables covering autonomy scope, accessible data/systems, human approvals, and cost/rate limits.

    That all Commonwealth tendering for the acquisition of AI capabilities include in the requirements document a clearly articulated set of AI harness capability parameters and deliverables specific to the 24 Agent Control Standard (n 2), 'About'. 25 National Institute of Standards and Technology, SP 800-53 Control Overlays for Securing AI Systems (COSAiS) (concept paper 2025; overlays in development), 'Using AI agent systems' use cases. https://csrc.nist.gov/projects/cosais September 23, 2026 16 THE LAYER ABOVE THE MODEL acquiring entity's mission. These should cover the permitted scope of autonomous action, the data and systems the agent may access, the actions that require human approval, and the applicable cost and rate limits.

  8. Resource the Australian AI Safety Institute to develop harness-level assurance and evaluation methods, so safety claims are warranted for the system as deployed rather than the model in isolation.

    That the Australian AI Safety Institute be resourced to develop harness-level assurance and evaluation methods, so that any claim that an AI system is safe, compliant or fit for purpose is warranted for the system as it is actually deployed rather than for the model in isolation.

  9. Have the AI Safety Institute draw on international work, including NIST’s AI security overlays, to develop Australian assurance methods alongside that effort.

    The Institute should draw on the international work now under way, including the control overlays for securing AI systems that NIST is developing26, so that Australian assurance methods are developed with, rather than apart from, that effort.

  10. Legislate tamper-evident harness records of model context, tools, approvals and actions in Commonwealth and nationally significant private deployments; require serious incident reports, including containment failures and unauthorised actions, to a designated authority.

    That the AI agent register requirements introduced into the Information Security Manual in September 2026 be generalised into a legislative harness telemetry requirement, so that AI harnesses used in Commonwealth systems, and in nationally significant private-sector deployments, are required to maintain tamper-evident records of the context supplied to the model, the tools called, the approvals given and the actions taken, and to report serious incidents, including containment failures and unauthorised actions, to a designated authority.

  11. Invest in and back sovereign AI harness capability, using acquisition practices to shape market parameters, with reference harness architectures and reusable patterns for public-sector and small-business use as targets.

    That the Government explicitly invest in and back sovereign AI harness capability development and deployment, using its acquisition practices to shape market parameters in parallel to achieving AI safety and security objectives. The targets for such investments should comprise reference harness architectures, reusable harness patterns suitable for public-sector and small-business use,

  12. Require model portability independent of any single overseas provider, and a domestic workforce program for designing, operating and assuring harnesses.

    requirement for model portability so that governance and continuity do not depend on any single overseas model provider, and a skills program to build a domestic workforce able to design, operate and assure harnesses.

  13. Align Australian harness standards and assurance with NIST’s AI Risk Management Framework, AS ISO/IEC 42001:2023, OWASP’s Agent Control Standard and NIST single-/multi-agent security overlays.

    That the Australian Standard and the associated assurance methods be aligned with the international work already forming at this layer. At the governance level, that alignment includes the United States NIST AI Risk Management Framework28 and ISO/IEC 42001, already adopted domestically as AS ISO/IEC 42001:202329, which set how AI risk is managed across an organisation, and for which a harness profile of the Australian Standard is the natural home. At the harness and agent-control level, where the most directly comparable work now sits, it includes the Agent Control Standard adopted into the OWASP GenAI Security Project in September 202630 and the control overlays for securing AI systems that NIST is developing for single-agent and multi-agent use31.

  14. Examine how assurance technology (AssureTech) and its RegTech subset can be applied to AI governance, and consider it a deliberate focus of national competitive advantage in planning the forward AI regulatory regime.

    That the Committee recommend the Australian Government examine how assurance technology (AssureTech), and its established regulatory-technology (RegTech) subset, can be applied to the governance of AI, treating the AI harness as assurance technology for AI, and consider assurance technology as a deliberate focus of national competitive advantage in planning the forward regulatory regime for AI.

  15. Name the AI harness explicitly within the scope of the Australian Standards for AI framework announced on 15 July 2026 and intended for legislation in early 2027.

    that the AI harness be named explicitly within the scope of the Australian Standards for AI framework announced by the Prime Minister on 15 July 2026 and intended for legislation in early 2027

Read the original

In some quotes a program corrected obvious text-recognition errors for display.

Open in the interactive guide