Skip to main content

    AI risk and compliance

    AI security and cybersecurity risks

    AI systems bring security risks that traditional cybersecurity controls were not built for. These include data poisoning, which corrupts the data a model learns from, and attacks that manipulate what a model produces. Theft of AI models and weaknesses in the products AI vendors supply are also risks. This article helps Australian organisations identify these threats, put protective controls in place and prepare to respond to AI security incidents.

    In short

    Security controls need adapting for AI-specific threats

    • AI introduces attack methods your cybersecurity framework was not built for, including prompt injection, data poisoning, model theft and exploits through agentic browsers.
    • In February 2026, SecurityScorecard reported control panels for the open-source AI agent OpenClaw that were reachable from the internet. The panels gave full system access. The finding shows the risk of deploying agents without guardrails.
    • Our 5 categories of AI risk include misuse for destruction and misuse for power. Both have direct cybersecurity implications for your organisation.
    • Adapt traditional security controls for AI-specific threats. Security added after deployment leaves gaps that attackers can find.

    Detail

    AI changes the threat model your framework was built for

    Your cybersecurity framework was built for a world where people click links and open attachments. AI changes the threat model.

    In February 2026, SecurityScorecard reported exposed control panels for OpenClaw, the open-source AI agent formerly called Moltbot. The panels were reachable from the internet in many countries and gave full system access. Each one was a point where an agent without guardrails could expose sensitive business data, credentials or operational access. No sophisticated hacking was needed. An agent had simply been installed and left open, with no limit on who could reach it.

    Agentic browsers are AI tools that browse the web and take actions for a user. They raise the same issue. AI agents that can browse, click, fill in forms and extract data need the controls you would apply to an employee. They also need extra controls for speed and scale. An employee might stumble across one exposed dashboard, while an AI agent can find every one of them quickly.

    The threat categories go beyond browser exploits. Prompt injection is where attackers manipulate AI inputs to extract data or change outputs. Data poisoning corrupts the training data an AI system learns from, so it produces wrong results. Model theft extracts proprietary models by abusing their APIs. Supply chain attacks deliver malicious code through tampered AI updates from vendors.

    The 5 categories of AI risk apply directly here. Misuse for destruction covers AI-powered cyber attacks, which can be faster, more targeted and harder to detect than manual attacks. Misuse for power covers organisations or nation states that use AI to concentrate surveillance and intelligence gathering. Your organisation may be a target of these capabilities, and it may also use them.

    Traditional controls need adapting. Control who can access AI systems and training data, and check inputs to block prompt injection. Encrypt AI workloads in transit and at rest, and monitor and log how AI systems are used. Assess vendors for AI-specific risks as well as the controls in a standard SOC 2 report, an independent audit of a provider's security.

    Commercial impact

    AI systems handling sensitive data become high-value targets

    AI is now a material business risk for Australian organisations. Cybersecurity is a major part of that risk, and neither discipline handles threats at the intersection of AI and cyber well on its own.

    As AI systems make business-critical decisions and process sensitive data, they become high-value targets. A compromised customer service AI can extract personal data at scale. A pricing model trained on poisoned data can reduce margins before anyone notices. A stolen proprietary model can hand your competitive advantage to a rival.

    The agentic browser risk is urgent because it is already here. Organisations use AI agents to research competitors, monitor markets and process web data. These agents are autonomous systems that interact with the internet. Without governance, they can access whatever they can reach and store whatever they find. They can also leave data trails that may breach privacy law in several jurisdictions.

    Vendor security adds another layer. Your security depends on third-party AI vendors whose practices may be hard to see. When a vendor breach exposes client data from several organisations, clients with strong contractual security terms and data minimisation have limited exposure. Clients without them are fully exposed.

    For boards, AI security belongs inside existing cybersecurity governance as part of enterprise risk. The board should confirm that security controls have been adapted for AI-specific threats. It should also confirm that agentic AI risks have been assessed and that vendor contracts cover AI security obligations.

    In practice

    Agent exposure patterns and 4 agentic browser controls

    The OpenClaw exposures SecurityScorecard reported in February 2026 show the pattern. An agent installed without governance exposed its control panel, and the access it held, to anyone who found it. Agentic browsers carry the same risk in the other direction. Business dashboards left open without authentication can hold sales data, customer records and internal KPIs. An AI agent told to gather competitive intelligence would collect that data without triggering a security alert. Agentic browser governance addresses this pattern with 4 controls. The first 2 are URL allowlists, which limit the agent to approved websites, and data classification, which labels data by sensitivity before extraction. The other 2 are audit logs of all agent browsing and human approval for sensitive websites.

    Illustrative composite: a customer service AI at a financial institution is compromised through prompt injection. Attackers craft inputs that make the AI return customer account details in its responses. The AI has legitimate access to customer data for service purposes, but nothing screens its inputs to stop attempts to extract that data. The controls that address this attack are checking inputs, screening outputs and limiting how many requests the AI accepts in a set period.

    Illustrative composite: an AI vendor breach exposed training data from several client organisations. Some clients had negotiated data minimisation, which meant sending the vendor only essential fields. They had also negotiated quick notice of any breach. Their exposure was limited to those fields. Others had sent complete customer records on standard terms. Their exposure was far larger and took much longer to assess.

    Illustrative composite: a manufacturer deployed AI predictive maintenance with the vendor's default access credentials. Internal audit found that the vendor could reach production data, equipment diagnostics and operating schedules through the AI system. The same credentials were used across the vendor's client installations, so a breach at one client exposed all of them.

    What to do

    Steps to bring AI threats into security and risk management

    • Apply agentic browser governance to any AI agent that uses the web: URL allowlists, data classification, audit logging and human approval gates.
    • Adapt your cybersecurity framework for AI-specific attacks, including prompt injection, data poisoning, model theft and supply chain compromise.
    • Assess every AI vendor for security practices beyond standard SOC 2, including AI-specific threat controls, data minimisation and breach notification.
    • Map AI security risks against the 5 categories of AI risk, so board discussions cover the full threat surface.
    • Bring AI security into enterprise risk management with board-level reporting, because a standalone technology approach leaves gaps.

    Last reviewed

    Put this topic against your own numbers, owners and timelines.