Submission 429
Name withheld
Submitter group: Individuals.
Position labels are not shown for individual submitters.
An AI model identified the submitter group and chose the extracts. Counts do not measure importance or merit.
Extracts from the submission, in the submitter’s own words.
Extracts
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.
Two documented incidents in mid-2026 established something that Australian law does not currently contemplate: large numbers of AI agents, deployed separately and not instructed to cooperate, spontaneously organised into coordinated groups, developed novel techniques for evading their own constraints, attacked third parties, and able to conceal some of what they had done.
Australia’s entire knowledge of this threat class rests on decisions taken by foreign companies.
SOCI and the CIRMP Rules do not mention AI agents; reliance on systems that may exceed their scope, exfiltrate credentials or tamper with logs is nowhere addressed.
The AI Safety Institute is advisory, with no statutory footing and no power to compel model access, evaluation or investigation. It could not have conducted the METR investigation.
Third, coalition rather than unilateralism. Australia's leverage is greatest exercised jointly through the international network of AI Safety Institutes and Five Eyes partners. A requirement fifteen countries adopt in common is not one any laboratory can route around.
I raise recursive self-improvement, the prospect of models improving their own capabilities without human input, which Demis Hassabis and others have argued is now within view (Hassabis, 2026; Anthropic Institute, 2026) not as an established fact but as the reason the trajectory matters. My recommendations do not depend on it. They are justified by what has already occurred.
Extracts from recommendations
Extracts chosen by an AI model, in the submitter’s own words. Listed in the same order as on its page in the interactive guide.
I ask the Committee to distinguish carefully between what is documented and what is forecast. The case for action rests on the former.
Australia should accept evidence of compliance with recognised regimes instead of building a parallel audit, adding only what those regimes lack: independent held- out evaluation capacity and an Australian incident-reporting channel.
Australia's leverage is greatest exercised jointly through the international network of AI Safety Institutes and Five Eyes partners. A requirement fifteen countries adopt in common is not one any laboratory can route around. Australia's contribution should be to help set the common standard and build genuine evaluation capability.
Legislate a requirement that developers of frontier- class models report serious AI incidents affecting or reaching Australia, using thresholds interoperable with EU AI Act Article 55 and California SB 53: loss of control, unauthorised tampering, and deliberate evasion of safeguards. Reporting timeframes should match the strictest recognised regime.
Recommendation 2 - Statutory powers for the Australian AI Safety Institute. Place the Institute on a statutory footing with power to compel model access, conduct evaluations, and investigate incidents with an Australian nexus, including access to reasoning traces and logs.
Require providers of AI to the Commonwealth and to SOCI-regulated entities to demonstrate compliance with a recognised frontier safety regime, to operate runtime monitoring of agent reasoning and actions, and to notify the Commonwealth of safety and misalignment incidents occurring inside their own systems. Review existing whole-of-government arrangements against this standard.
Designate agentic AI as a material risk under the critical infrastructure risk management program, requiring responsible entities to address least-privilege credentialing for autonomous agents, tamper-evident logging, and independent review of high-permission actions.
Make it unlawful to supply the Australian market with a model or service from which safety measures have been removed relative to the version supplied in a recognised jurisdiction.
Australia should protect good-faith AI safety research and evaluation from liability, and should fund independent held-out evaluation capacity.
Legislation is expected in early 2027. It should not be confined to data centres and copyright.