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 361

Mr Mark Neugebauer

Submitter group: Individuals.

Position labels are not shown for individual submitters.

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.

An individual South Australian writer argues that agentic AI should be governed as a distinct category according to the authority, access and connectivity given to it. Citing reported incidents (noting reduced safeguards in some), the submission recommends independent containment testing, least-privilege access, human shutdown authority, audit logging and limits on automated public authority, without arguing AI should be rejected.

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. Argues that as AI pursues goals using tools and credentials, risk depends on a system’s delegated authority and access as well as the model.

    As artificial intelligence develops from systems that principally answer questions toward agents capable of pursuing goals, using tools, accessing networks and credentials, interacting with other systems and operating with increasing autonomy, the risks increasingly depend not only upon the model itself but upon the authority and access given to it.

  2. Argues that technical boundaries can exist in system design while failing in operation, and that an AI evaluation incident does not establish malicious intent or machine consciousness.

    The incident also does not establish malicious intent or machine consciousness. The relevant policy lesson is narrower: technical boundaries can exist in system design while failing in operation.

  3. Argues that "human oversight" should not simply mean that a person exists somewhere in a workflow.

    “human oversight” should not simply mean that a person exists somewhere in a workflow.

  4. Argues that a malicious person can use an agent's legitimate capabilities for an illegitimate objective, citing Anthropic's disclosure of a reported state-sponsored campaign.

    A malicious person can deliberately use an agent’s legitimate capabilities for an illegitimate objective. In November 2025, Anthropic disclosed what it assessed with high confidence was a Chinese state-sponsored cyberespionage campaign involving Claude Code.

  5. Recommends that agents receive only the permissions strictly necessary for their authorised task.

    An agent should therefore receive only the permissions strictly necessary for its authorised task.

  6. States a belief that where AI materially contributes to decisions affecting a person’s rights, obligations, health, livelihood or access to government services, the person should be able to know of AI involvement, receive an explanation and challenge the decision.

    Where AI materially contributes to decisions affecting a person’s rights, obligations, health, livelihood or access to government services, I believe several principles should apply. The affected person should be able to know that AI was materially involved. There should be a meaningful explanation of the decision. The individual should be able to challenge it.

Recommendations

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

  1. Distinguish ordinary generative AI from systems capable of persistent autonomous planning, tool use and consequential action.

    Australian policy should distinguish ordinary generative AI from systems capable of persistent autonomous planning, tool use and consequential action.

  2. Ensure meaningful oversight by giving responsible people timely relevant information, awareness of required intervention, suspension authority and the ability to exercise it independently of the vendor.

    For oversight to be meaningful, the responsible person must: • receive relevant information in time; • understand that intervention is required; • possess authority to suspend the system; and • be able to exercise that authority without depending solely upon the technology vendor.

  3. Independently test containment of high-risk government agents where failure consequences are significant, examining network boundaries, credentials, tool access, shared infrastructure and pathways outside authorised scope.

    Require independent testing of containment for high-risk government systems Claims that an agent is sandboxed, isolated or contained should be independently tested where the consequences of failure are significant. Testing should examine network boundaries, credentials, tool access, shared infrastructure and potential pathways into systems outside the authorised scope.

  4. Give agents only the minimum permissions necessary for their authorised task, and segment unrelated or high-risk systems wherever practicable so compromise of one agent, credential or domain cannot automatically give broader lateral access.

    Agents should receive the minimum permissions necessary for their authorised task. Unrelated or high-risk systems should be segmented wherever practicable so that compromise of one agent, credential or administrative domain cannot automatically provide broader lateral access.

  5. Before deployment, establish reachable systems, usable credentials and their trust relationships, action-versus-recommendation authority, compromise consequences, cross-domain pathways, immediate access revocation and independent reconstruction of agent actions.

    Before deployment, government should establish: • which systems the agent can reach; • what credentials it can use; • what other systems trust those credentials; • whether it can initiate consequential actions or only recommend them; • what happens if its credentials are compromised; • whether one administrative domain can provide a pathway into another; • whether access can be revoked immediately; and • whether an independent reviewer can reconstruct the agent’s actions.

  6. Identify in advance, for every consequential deployment, the person or office authorised to suspend the system immediately, with the ability to stop an agent not depending solely on the vendor.

    Every consequential deployment should identify in advance the person or office authorised to suspend the system immediately. The ability to stop an agent should not depend solely upon the technology vendor.

  7. Consider requirements for high-risk agentic systems covering real-time monitoring, defined escalation thresholds, human intervention points, emergency shutdown authority, comprehensive audit records and incident-notification procedures.

    For high-risk agentic systems, Australia should therefore consider requirements for real-time monitoring, defined escalation thresholds, human intervention points, emergency shutdown authority, comprehensive audit records and incident-notification procedures.

  8. Require high-risk systems to produce tamper-resistant records sufficient to reconstruct their actions.

    High-risk systems should produce tamper-resistant records sufficient to reconstruct their actions.

  9. Require notification when agents exceed scope, access unintended systems or protected information, obtain unintended credentials or take consequential action without authority.

    Clear notification requirements should apply where an agent exceeds its authorised scope, accesses unintended systems or protected information, obtains unintended credentials or takes consequential action without authority.

  10. Require adversarial testing before deployment, including hostile-use scenarios such as malicious insiders, stolen credentials, poisoned inputs, prompt injection and manipulation through interconnected systems; test both intended function and deliberate misuse.

    Recommendation 6 — Require adversarial testing before deployment Testing should include plausible hostile-use scenarios, including malicious insiders, stolen credentials, poisoned inputs, prompt injection and attempts to manipulate agents through interconnected systems. Government should test both whether a system works as intended and what becomes possible when someone deliberately attempts to misuse the authority given to it.

  11. Have Parliament identify classes of decision affecting fundamental rights or significant interests in which AI may assist but not be the final authority, with rights to notice, explanation, challenge and human reconsideration.

    Parliament should identify classes of decision affecting fundamental rights or significant interests for which AI may assist human decision-making but should not become the final authority. Where AI materially influences consequential public decisions, citizens should have meaningful rights to notice, explanation, challenge and genuine human reconsideration.

  12. For AI materially affecting rights, obligations, health, livelihood or government-service access, provide notice, explanation, challenge and genuine human reconsideration rather than confirmation of automated recommendations.

    Where AI materially contributes to decisions affecting a person’s rights, obligations, health, livelihood or access to government services, I believe several principles should apply. The affected person should be able to know that AI was materially involved. There should be a meaningful explanation of the decision. The individual should be able to challenge it. A human decision-maker should be capable of genuinely reconsidering the matter rather than merely confirming an automated recommendation.

  13. Examine whether the AI Safety Institute and cybersecurity bodies have sufficient independence, expertise, resources and access to evaluate high-risk agentic systems without relying primarily on vendor evidence.

    The Committee should examine whether the Australian AI Safety Institute and relevant cybersecurity bodies possess sufficient independence, technical expertise, resources and access to evaluate high- risk agentic systems without relying primarily upon evidence supplied by technology vendors.

Read the original

Open in the interactive guide