Submission 45
Palo Alto Networks
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.
Palo Alto Networks, a cybersecurity company, argues that AI is both a major new attack surface and a powerful defensive capability. It describes AI supply-chain and critical-infrastructure risks and its own frontier-model defence work, and asks for secure-by-design practices, managed defender access to frontier models and agent governance in the public sector.
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 AI ecosystems introduce attack surfaces that traditional network and application security tools were not designed to address.
AI models, applications, agents, and datasets that introduces attack surfaces that traditional network and application security tools were never designed to address, including prompt injection, sensitive data leakage, model theft, training data poisoning, insecure plugin design, and agentic threats such as tool misuse, identity impersonation, and memory manipulation.
Argues that modern AI systems incorporate publicly sourced models, agent skills, third-party plugins and datasets of uncertain provenance.
Modern AI systems are rarely built from scratch; they are assembled from models sourced from public repositories, agent "skills" from emerging marketplaces, third-party plugins, and datasets of uncertain provenance.
Says industry research indicates many vulnerabilities are weaponised within 24 hours of disclosure while enterprises average around 55 days to deploy patches.
Industry research indicates that a substantial share of vulnerabilities are now weaponised within 24 hours of public disclosure, while the industry-average time for an enterprise to test and deploy a software patch remains around 55 days.
Argues that regulators, the AI Safety Institute and security agencies need early, well-governed frontier-model access for defence, alongside infrastructure operators and security providers.
the pace at which Australian regulators, the Australian AI Safety Institute, and the national intelligence and security community can identify and respond to frontier-AI-enabled threats will depend materially on whether they, and the critical infrastructure operators and security providers they work with, can obtain comparably early and well-governed access to frontier model capabilities for defensive purposes
Argues that a compromised AI model or agent skill in a government or critical infrastructure environment could be a channel for espionage, data exfiltration or manipulation of automated decisions.
A compromised or maliciously tampered AI model or agent skill deployed inside a government or critical infrastructure environment is not merely a software defect; it is a potential channel for espionage, data exfiltration, or the covert manipulation of automated decision-making.
Recommendations
Recommendations are shown beside the submission’s own words. Listed in the same order as on its page in the interactive guide.
Recommend pre-production validation of AI models and agent skills for malicious code, unsafe configurations and licensing issues as best practice for Australian organisations, including government agencies.
we recommend the Committee reflect the following defensive capabilities that map directly onto the AI attack lifecycle as best practices for Australian organisations, including government agencies, to adopt widely: 1. validating AI models and agent skills for malicious code, unsafe configurations, and licensing issues before they are deployed into production;
Continuously red-team AI applications and agents against recognised frameworks, including OWASP LLM Top 10, MITRE ATLAS and NIST AI Risk Management Framework.
continuously red-teaming AI applications and agents against recognised frameworks such as the OWASP LLM Top 10, MITRE ATLAS and the NIST AI Risk Management Framework to surface vulnerabilities before adversaries find them;
Inspect prompts, responses and data flows in real time to detect and block prompt injection, sensitive data leakage and malicious code.
inspecting prompts, responses, and data flows in real time to detect and block threats such as prompt injection, sensitive data leakage, and malicious code;
Secure autonomous agents against identity impersonation, credential leakage and manipulation of memory or tool use.
securing autonomous agents against threats specific to agentic systems, including identity impersonation, credential leakage and manipulation of an agent's memory or tool use;
Maintain ongoing visibility into AI ecosystem misconfigurations, excessive permissions and sensitive data exposure.
maintaining ongoing visibility into misconfigurations, excessive permissions, and sensitive data exposure across the AI ecosystem.
Reflect AI vulnerability discovery, including multi-agent research, with early coordinated disclosure among cloud, software and operational-technology vendors as sector best practice, enabling network shielding within hours.
We recommend the Committee reflect, as a best practice for the sector, the pairing of AI-driven vulnerability discovery, including through multi-agent vulnerability research systems, with early, responsibly coordinated disclosure across a broad industry coalition spanning cloud, software and operational technology vendors, so that protective network-level shielding can be deployed within hours rather than the standard weeks-to-months patch cycle.
Encourage coalition-based, AI-enabled defence more broadly.
We highlight this only to illustrate the category of coalition-based, AI-enabled defence that we believe the Committee should encourage more broadly.
Resource the AI Safety Institute and national security bodies to facilitate secure, managed access to frontier model capabilities for legitimate defensive cyber purposes, including for critical infrastructure operators and their security vendors, not only to evaluate risk.
The Australian AI Safety Institute and relevant national security bodies should be resourced not only to evaluate and monitor frontier AI systems, but to actively facilitate secure, managed access to frontier model capabilities for legitimate defensive cyber security purposes, including for critical infrastructure operators and the security vendors that support them.
Regulators and the intelligence and security community should evaluate their detection and response tooling on the assumption that adversaries may compress attack timelines to minutes, with implications for staffing, tooling procurement and government-industry information sharing.
Regulators and the intelligence and security community should evaluate their own detection and response tooling against the assumption that adversaries may compress attack timelines to minutes, not days. This has direct implications for staffing, tooling procurement, and information-sharing arrangements between government and industry, given that most Australian critical infrastructure is privately owned and operated.
Recommend public-sector autonomous-agent safeguards: inventory, verifiable identity, protocol security, continuous monitoring, human approval for high-impact or irreversible actions, autonomy-scaled oversight, trusted-registry verification and auditable fleet visibility.
we encourage the Committee to recommend that Australian public sector agencies deploying autonomous AI agents work toward adopting: a live, authoritative inventory of every deployed agent; a unique, cryptographically verifiable identity for each agent; dedicated security controls for the protocols agents use to connect to tools and to each other; continuous behavioural monitoring rather than one-off pre-deployment testing; human-in-the-loop approval for high-impact or irreversible agent actions; a risk-tiering process that scales deployment and oversight to an agent's level of autonomy; verification of agents' tools, skills, and models against a trusted registry before use; and centralised, auditable visibility across the entire agent fleet.
Recommend that regulators, intelligence and security agencies and the AI Safety Institute deepen operational integration with international vulnerability coalitions and defensive AI initiatives, alongside allied governments and industry.
We encourage the Committee to recommend that Australian regulators, intelligence and security agencies, and the Australian AI Safety Institute (AISI) deepen their operational integration with international vulnerability coalitions and defensive AI initiatives, alongside allied governments and industry, so that Australian critical infrastructure benefits from the same early-warning and rapid protection mechanisms available elsewhere.