Submission 276
MUZZ AI Pty Ltd
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.
MUZZ AI describes its controlled-cooperation architecture for sharing permitted information while preventing automatic transfer of consequential judgment between AI systems. It seeks independent government evaluation of its own technology. Supplements report laboratory containment results, retain the need for independent adversarial and partner testing, and complain about government and committee non-engagement.
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.
States its central principle: useful information may pass between AI systems, but one system's consequential judgment should not automatically become another's decision contribution.
useful permitted information may pass between systems, but one system's consequential judgment or decision authority should not automatically become another system's decision contribution.
Cites the AI Safety Institute's August 2026 report as identifying 'propagation and contagion' as a distinct failure class in multi-agent systems.
The Australian AI Safety Institute's August 2026 report is highly relevant. It begins from the premise that a system composed of individually safe and reliable agents is not necessarily a safe and reliable multi-agent system. New failures can emerge from the interactions themselves. The report identifies 'propagation and contagion' as a distinct failure class.
Argues that existing cyber security controls are necessary but address a different question from how much decision authority one connected AI should have over another.
Those controls remain necessary. MUZZ does not replace them. A conventional firewall or access-control system asks whether a communication or actor is permitted. MUZZ addresses an additional question after legitimate systems are permitted to cooperate: what authority should one participating AI have over the consequential decision-making of another?
Reports internal tests including 4,096 attempted transfers of a non-permitted type with none succeeding, and states these are not independent government certification.
a focused test involving 4,096 attempts to transfer a non-permitted type across the tested cooperation boundary with zero successful transfers. These are MUZZ's internal results, not independent government certification.
Argues that if financial-services systems inherit one another's conclusions, false AI consensus can result.
In financial services, multiple specialist systems may contribute to fraud detection, claims, underwriting, pricing, identity, transaction monitoring, customer service and cyber security. If one system's conclusion is simply inherited by others, an organisation can create false AI consensus: several systems appear to agree when the later systems are actually responding to the first system's judgment.
Says it believes it has solved containment of malicious, faulty or unauthorised commands, judgments and influence between cooperating AI systems, supported by laboratory evidence.
We believe we have solved the central problem of containing malicious, faulty and unauthorised commands, judgments and influence between cooperating AI systems. We have substantial laboratory evidence supporting that belief.
Recommendations
Recommendations are shown beside the submission’s own words. Listed in the same order as on its page in the interactive guide.
Recommend technical evaluation of controlled-cooperation architectures that contain consequential judgment between independent AI systems.
MUZZ AI Pty Ltd respectfully asks the Committee to consider recommending: 1. technical evaluation of controlled-cooperation architectures that seek to contain consequential judgment between independent AI systems;
Promote engagement between the AI Safety Institute, cyber and technical agencies, and Australian developers of multi-agent containment mechanisms.
engagement between the Australian AI Safety Institute, relevant cyber-security and technical agencies, and Australian developers working on multi-agent containment mechanisms;
Consider inter-system decision authority as a distinct issue in government AI procurement and high-consequence AI governance.
consideration of inter-system decision authority as a distinct issue in government AI procurement and high- consequence AI governance; and
Where appropriate, arrange controlled evaluation of MUZZ by a suitable Australian Government technical body.
4. where appropriate, controlled evaluation of MUZZ by a suitable Australian Government technical body.
Evaluate MUZZ under controlled conditions: performance, containment while sharing information, receiver control, adversarial behaviour, remaining limitations and suitability for financial services, infrastructure, defence or other consequential uses.
We ask the Committee to recommend that an appropriate Australian Government technical capability be given the opportunity to examine the architecture under controlled conditions and determine: whether the technology performs as described; whether consequential judgment or decision authority from a sender can be contained while permitted useful information continues to pass; whether the receiver genuinely retains control over admission and its own consequential determination; how the architecture behaves when a participant is erroneous, malicious, compromised or attempts to transmit a prohibited type; what limitations, bypass paths or failure modes remain; and whether the capability could contribute to safer multi-system AI in financial services, critical infrastructure, defence or other high-consequence environments.
Encourage architectures that permit AI cooperation while constraining one system’s authority over another.
It should encourage architectures that allow systems to cooperate while constraining the authority that one system can exercise over another.
Arrange independent adversarial evaluation and testing in a real partner environment, recognising laboratory results do not establish elimination of every cyberattack, covert channel or AI failure.
These remain internal laboratory results. They are not presented as proof that every possible cyberattack, covert channel or AI failure has been eliminated. Independent adversarial evaluation and testing in a real partner environment remain necessary.
If the inquiry only identifies and discusses AI dangers, consider a separate inquiry into how those problems can be fixed, beginning with the submitted Australian technology and laboratory evidence.
Is this an inquiry genuinely interested in solving the dangers of artificial intelligence, or is it simply another taxpayer-funded discussion about how dangerous AI might become? If it is the latter, perhaps Australia needs a separate inquiry into how these problems can actually be fixed. We suspect that inquiry could be considerably shorter and considerably less expensive for Australian taxpayers. It could begin by examining the Australian-developed technology and laboratory evidence already placed before this committee.
Identify comparable demonstrated containment solutions presented by the major organisations and experts whose evidence the committee considers.
We challenge it to identify what comparable demonstrated containment solutions have been presented by the major organisations and experts whose evidence receives its attention.
Explain why Australian developers apparently need substantial financial resources or prestigious institutional connections for their achievements to be taken seriously.
And we challenge it to explain why Australian developers apparently need substantial financial resources or prestigious institutional connections before their achievements are taken seriously.