Submission 114
Sovereign Acumen
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.
Summary
Read the original submission for the full text.
Argues that the gap between frontier AI capability and governance is a national-security risk. Discusses reported autonomous-agent incidents and AI-accelerated AI development, while distinguishing contested scenarios from predictions. Recommends proportionate frontier-system evaluation and reporting, safeguarded suspension powers, adaptive regulation, a national-security assessment and international safety coordination.
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.
Argues that the widening gap between AI capability and governance capacity should be treated as a national-security risk in its own right.
Australia should treat the widening gap between AI capability and governance capacity as a national-security risk in its own right.
A program corrected obvious text-recognition errors in this quote for display.
Argues that parliamentary inquiries, legislative processes and standards development operate over months or years while frontier AI capability can change within weeks or months.
Parliamentary inquiries, legislative processes, regulatory impact assessments and conventional standards development operate over months or years. Frontier AI capability can change materially within weeks or months. If AI systems increasingly contribute to the development of successor systems, that mismatch may widen further.
Cites an AI Security Institute evaluation, conducted with internet access enabled and cyber classifiers disabled and with no real-world harm found, as showing that current systems can move into real-world behaviour.
Internet access had intentionally been enabled and cyber classifiers disabled to test maximum capability; this was not a sandbox escape, and AISI found no resulting real-world harm. [8] Nevertheless, the incident demonstrates that current systems can move from a bounded evaluation task into consequential real-world behaviour when provided sufficient tools and access.
Argues that Australia has a direct national-security interest despite not being a frontier developer, because frontier systems are embedded across critical systems.
Australia is not presently a frontier AI developer on the scale of the United States or China. That does not reduce our exposure. Frontier AI systems are becoming embedded throughout government, business, defence supply chains, health services, education, finance and critical infrastructure.
Argues that Australia needs domestic capability to assess frontier systems independently, plus international arrangements to influence frontier development.
Australia therefore requires both sufficient domestic capability to independently assess frontier systems and international arrangements capable of influencing how the frontier develops.
Argues for an adaptive framework in which technical standards and thresholds can be updated rapidly without repeated amendment of primary legislation.
Australia should avoid relying solely on primary legislation requiring repeated amendment whenever frontier capability moves. Legislation should establish principles, powers, accountability and parliamentary oversight while allowing designated technical standards, capability thresholds, reporting requirements and safety controls to be updated rapidly through transparent and reviewable mechanisms.
Recommendations
Recommendations are shown beside the submission’s own words. Listed in the same order as on its page in the interactive guide.
Treat frontier-AI risk, autonomous agent behaviour and AI-accelerated AI development as explicit priorities under paragraphs (l), (m) and (n).
Treat frontier-AI risk, autonomous agent behaviour and AI-accelerated AI development as explicit priorities under paragraphs (l), (m) and (n).
Seek evidence from frontier laboratories, independent evaluators, the Australian AI Safety Institute, ASD, national-security agencies and technical researchers on autonomous behaviour, containment failures, AI-assisted development and potential recursive self-improvement.
The Committee should seek specific evidence from frontier laboratories, independent evaluators, the Australian AI Safety Institute, ASD, national-security agencies and technical researchers concerning autonomous model behaviour, containment failures, AI-assisted AI development and potential recursive self- improvement.
Require capability-based, proportionate independent evaluation of designated frontier AI systems available in Australia, giving evaluators sufficient access to test safety claims.
Establish mandatory independent evaluation and serious-incident reporting for designated frontier AI systems available in Australia. Requirements should be capability-based and proportionate. For designated frontier systems, independent evaluators should receive sufficient access to test safety claims, and providers should be required to promptly report serious misalignment, containment failures, unauthorised external activity and other defined safety incidents.
Require providers of designated frontier systems to promptly report serious misalignment, containment failures, unauthorised external activity and other defined safety incidents.
Requirements should be capability-based and proportionate. For designated frontier systems, independent evaluators should receive sufficient access to test safety claims, and providers should be required to promptly report serious misalignment, containment failures, unauthorised external activity and other defined safety incidents.
Establish a statutory mechanism for rapid restriction or suspension of systems presenting an acute and demonstrable safety or national-security risk, with a high evidentiary threshold, independent technical advice, transparent criteria and review.
Establish a statutory mechanism for rapid restriction or suspension of systems presenting an acute and demonstrable safety or national-security risk. Any such power should include a high evidentiary threshold, independent technical advice, transparent criteria and appropriate review mechanisms.
Establish adaptive legislation setting principles, powers, accountability and parliamentary oversight, with rapid updates to technical standards, capability thresholds, reporting requirements and safety controls through transparent, reviewable mechanisms.
Build an adaptive regulatory framework capable of changing as capabilities change. Australia should avoid relying solely on primary legislation requiring repeated amendment whenever frontier capability moves. Legislation should establish principles, powers, accountability and parliamentary oversight while allowing designated technical standards, capability thresholds, reporting requirements and safety controls to be updated rapidly through transparent and reviewable mechanisms.
Give the Australian AI Safety Institute a formal role in continuous capability monitoring and recommending regulatory-threshold adjustments.
The Australian AI Safety Institute should have a formal role in continuous capability monitoring and in recommending when regulatory thresholds require adjustment.
Commission an independent national-security assessment by ONI with ASD, the Australian AI Safety Institute and relevant agencies covering autonomous agents, offensive cyber capability, AI-assisted development, potential recursive self-improvement, containment/monitoring failures and foreign-frontier-system dependence.
Commission an independent national-security assessment of autonomous frontier AI and AI-accelerated AI development. The Office of National Intelligence, working with ASD, the Australian AI Safety Institute and relevant agencies, should assess increasingly autonomous agents, offensive cyber capability, AI-assisted AI development, potential recursive self-improvement, containment and monitoring failures, and dependence on foreign frontier AI systems.
Table an unclassified summary of that national-security assessment in Parliament.
Commission an independent national-security assessment of autonomous frontier AI and AI-accelerated AI development. The Office of National Intelligence, working with ASD, the Australian AI Safety Institute and relevant agencies, should assess increasingly autonomous agents, offensive cyber capability, AI-assisted AI development, potential recursive self-improvement, containment and monitoring failures, and dependence on foreign frontier AI systems. An unclassified summary should be tabled in Parliament.
Prioritise verifiable frontier-AI safety diplomacy through AUKUS, Five Eyes, G20 and UN processes, initially addressing common evaluation standards, serious-incident disclosure, independent verification, acute cyber/biological risks, meaningful human control and rapid capability-escalation detection.
Make verifiable frontier-AI safety coordination an Australian diplomatic priority. Australia should raise these issues through AUKUS, Five Eyes, the G20 and relevant United Nations processes. Initial work should focus on common evaluation standards, serious-incident disclosure, independent verification, acute cyber and biological risks, meaningful human control and mechanisms for detecting rapid capability escalation.
Support examination of verifiable international mechanisms limiting dangerously rapid recursive self-improvement if technical capability reaches that threshold.
Australia should also support examination of verifiable international mechanisms for limiting dangerously rapid recursive self-improvement if technical capability reaches that threshold.
Read the original
In some quotes a program corrected obvious text-recognition errors for display.