Radware’s latest threat intelligence report, The Internet of Agents: The Next Threat Surface, marks a turning point in how organisations and regulators must view artificial intelligence. AI agents are no longer just tools that execute commands; they are becoming autonomous actors capable of reasoning, decision-making, and inter-agent collaboration. For policymakers and risk managers in Australia and worldwide, this evolution brings pressing questions of accountability, compliance, and regulation.
Defining the New Agentic Layer
Unlike conventional APIs or software functions, agentic AI operates at a higher order of complexity. Powered by large language models (LLMs), these agents can independently use tools, access systems, and exchange context with one another via protocols such as the Model Context Protocol (MCP) and Agent-to-Agent (A2A) communication. This makes them powerful, but also unpredictable. Left unchecked, AI agents can escalate privileges, cross trust boundaries, and act in ways that are difficult for human overseers to anticipate or constrain.
Policy and Regulatory Considerations
Agent Identity and Accountability
One of the most pressing governance challenges is clarifying accountability. If an AI agent carries out an unauthorised action, who bears responsibility—the developer, the platform provider, or the organisation deploying it? Clear accountability frameworks will be required. In practice, this means treating agents as privileged identities, with audit logs, traceability, and transparent decision-making trails built into their design and use.
Regulating Protocols
Emerging standards like MCP and A2A should not be overlooked. While they enable interoperability and automation, they also create potential risk points. Regulators may need to consider whether these protocols require oversight, particularly in industries handling sensitive data or operating under critical infrastructure mandates.
Addressing Malicious AI Platforms
The rise of subscription-based services like XanthoroxAI illustrates how attack toolchains are being commoditised. Just as ransomware-as-a-service prompted regulatory scrutiny, policymakers may need to examine whether platforms that facilitate AI-driven cybercrime should be classified, restricted, or monitored more closely.
Vulnerability Disclosure and Response
The accelerating speed of exploit generation by LLMs is another regulatory flashpoint. With AI capable of rapidly transforming vulnerability descriptions into working exploits, current disclosure and patching practices may be too slow. Regulators could respond by tightening expectations around patching windows, revisiting responsible disclosure frameworks, and mandating faster vulnerability response cycles.
Vendor and Supply Chain Risk
For enterprises, third-party AI agents and tools pose a new dimension of supply chain risk. Vendor assessments can no longer stop at traditional security questionnaires; they must probe how agents are built, what safeguards exist against prompt injection, and whether resilience is embedded in their design. Transparency from vendors and enforceable security assurances will be key in mitigating cascading risks.
The Road Ahead
As AI agents continue to evolve, they will occupy an increasingly central role in both business operations and cyberattack strategies. The challenge for policymakers and governance leaders will be to balance innovation with accountability, ensuring that automation does not outpace the regulatory frameworks designed to keep society secure.
Radware’s report makes one point clear: the agentic AI era demands new thinking across technical, legal, and ethical domains. For risk managers, this means building governance practices that are as dynamic as the technologies they oversee. For policymakers, it means anticipating the consequences of AI-driven autonomy before adversaries exploit them.
