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 224

CI-ISAC Australia

Submitter group: Industry bodies.

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.

A cyber threat intelligence sharing organisation describes, from operational experience, how AI affects critical infrastructure. It argues that the existing SOCI framework, rather than a new AI-specific regime, should be extended to AI supply chain dependencies, with guidance suited to small operators, support for sector sharing networks, an AI Safety Institute-ASD channel and sovereign threat data.

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. Says that AI has helped adversaries quickly and cheaply, while defender productivity has improved slowly and mostly for large entities.

    Adversary productivity has improved immediately and cheaply. Defender productivity has improved slowly, unevenly, and mostly for large, well-resourced entities.(ToR l) CI-ISAC's quarterly threat reporting to members has tracked adversary use of AI maturing from experimental to operational. The practical effect of AI on Australian critical infrastructure to date has been asymmetric.

  2. Says unsanctioned use of public large language models by clinical and administrative staff is the most frequent issue raised with it, and is almost never malicious.

    Unsanctioned use of public large language models by clinical and administrative staff is the most frequent issue raised with us. It is almost never malicious: a staff member under time pressure pastes a discharge summary, patient correspondence or an incident report into a general-purpose assistant to have it summarised.

  3. Argues that most critical infrastructure operators acquire AI capability inside products they already buy rather than building AI systems.

    Most critical infrastructure operators are not building AI systems. They are acquiring AI capability inside products they already buy: electronic medical records, scheduling and triage tools, claims processing, customer service platforms, managed security services, and increasingly the security tooling itself.

  4. Argues that the existing SOCI framework can carry AI dependencies and that what is missing is authoritative guidance on treating them as supply chain hazards.

    We do not recommend a new AI-specific critical infrastructure regime. The existing SOCI framework is capable of carrying this. What is missing is authoritative guidance that tells responsible entities how to treat AI dependencies as material supply chain hazards, and what an adequate answer looks like.

  5. Argues that no routine channel connects AI model-level risk findings to those defending networks, or operator observations back.

    Together they leave a gap: there is no routine, timely channel that moves a finding about AI model capability or misuse into the hands of the people defending Australian hospitals, water utilities and payment systems, and no channel that moves operator observations back the other way.

  6. Argues that a narrow model trained on pooled Australian critical infrastructure telemetry would detect what globally trained tools miss.

    Our sectoral structures, regulatory environment, institutions and adversary targeting patterns differ from those reflected in globally trained detection models, so a narrow model trained on pooled Australian critical infrastructure telemetry would detect what global tooling misses.

Recommendations

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

  1. Treat AI dependencies as critical infrastructure supply-chain dependencies; coordinate AI regulation with SOCI Act reform and clarify existing risk obligations for models, inference providers and AI-enabled managed services relied on by responsible entities.

    Treat AI dependencies as critical infrastructure supply chain dependencies. Ensure that AI-related regulatory frameworks are developed in concert with the reform agenda of the Security of Critical Infrastructure Act 2018, and clarify how existing risk management obligations apply to AI models, inference providers and AI-enabled managed services relied on by responsible entities.

  2. Support sector-wide bill of materials capability as the means of making AI dependencies visible, extending established SBOM practice to AI components.

    Support sector-wide bill of materials capability as the practical means of making those dependencies visible, extending established SBOM practice to AI components.

  3. Accompany AI governance frameworks with operational guidance a small provider can act on without interpretation (approved-tool lists, a data classification rule, standard vendor questions), delivered through sector sharing networks.

    Pitch AI guidance at the altitude of the organisations expected to use it, and deliver it through sector networks. Governance frameworks should be accompanied by operational guidance a small provider can act on without interpretation: approved-tool lists, a memorable data classification rule, and a standard set of vendor questions.

  4. Fund sector-based sharing as national defensive infrastructure, implementing Recommendation 6(a) of the Slay Review of the SOCI Act on bi-directional sharing.

    Fund sector-based sharing as national defensive infrastructure. Implement Recommendation 6(a) of the Slay Review of the SOCI Act on bi-directional sharing, with sector sharing networks as the practical conduit between government intelligence holdings and operator visibility.

  5. Recognise validated, human-reviewed threat intelligence as a national capability.

    Recognise validated, human-reviewed threat intelligence as a national capability. AI-generated security content that is confidently wrong is now polluting the information environment defenders rely on, and validation at sector level is what keeps it out.

  6. Recognise Australian cyber threat data as a sovereign asset and create a clear legal pathway for it to be pooled, held onshore under trusted custodianship, and used to train narrow, domain-specific defensive models.

    Recognise Australian cyber threat data as a sovereign asset and create a clear legal pathway for it to be pooled, held onshore under trusted custodianship, and used to train narrow, domain-specific defensive models.

  7. Avoid framing security primarily as frontier-model risk; recognise immediate harms from ordinary crime industrialised with widely available tools.

    None of this requires frontier capability. It is being done with widely available tools. The Committee should be cautious about framing the security question as primarily one of frontier model risk. The immediate national harm is the industrialisation of ordinary crime against organisations that were already under-resourced.

  8. Provide small organisations operational support including approved tools, memorable data classification, egress visibility and vendor questions.

    What these organisations need is operational: an approved-tool list, a data classification rule simple enough to remember, egress visibility, and a short set of questions to put to vendors embedding AI in existing products.

  9. Establish standing operational liaison between the AI Safety Institute, ASD and accredited sector sharing networks, explicitly exchanging AI threat and capability findings at speed.

    We recommend a standing operational liaison between the AI Safety Institute, ASD and accredited sector sharing networks, with an explicit remit for exchanging AI-relevant threat and capability findings at speed.

  10. Ensure pooled threat data has a clear legal pathway across privacy, competition and security law, and is held onshore under trusted, member-accountable custodianship.

    Two conditions must be met. The data has to be pooled, which requires a legal pathway presently unclear across privacy, competition and security law: standing up our Finance Cyber Sharing Network required an ACCC authorisation application, which illustrates the friction involved in lawful collective defence. And it has to be held onshore under trusted, member-accountable custodianship, or operators will not contribute.

  11. Fund the baseline infrastructure needed for intelligence sharing, as well as the exchange itself.

    Sharing arrangements designed for a mature security operations capability will reach the organisations that least need help and miss those that most do. The baseline needs funding, not only the exchange.

  12. Direct National AI Plan adoption and uplift support through sector sharing networks, in operational rather than governance form.

    Sector sharing networks are the most efficient existing channel for uplift to these entities. We recommend that adoption and uplift support under the National AI Plan be directed through them, take operational rather than governance form,

  13. Recognise participation in sector sharing, in critical infrastructure risk management, as evidence of appropriate monitoring of the threat environment.

    participation in sector sharing be recognised in critical infrastructure risk management as evidence of appropriate monitoring of the threat environment.

Read the original

Open in the interactive guide