This article was co-authored by Dietrich Marquardt, Bradley Phillips and Damien Vickovich.

Key takeaways

  • The cybersecurity capability of next-generation frontier artificial intelligence (AI) models is increasing rapidly.  Large language models (LLM) such as Claude Mythos 5 and Fable 5 are enabling IT vulnerabilities to be discovered faster than ever before. The launch of these models prompted global concern, public calls to action, and regulatory responses, including from ASIC and APRA in Australia. Governments also began intervening in the release of the most capable models.
  • On 22 June 2026, each of the ‘Five Eyes’ cyber security agencies issued an urgent call to action that “the evolving landscape of AI is rapidly transforming cyber risk, and we must act swiftly to remain ahead”. The agencies noted that cyber risk is not just a technical issue, but requires a whole of organisation and whole of society response. They commented that boards and executives must act now.
  • The legal standard for boards has not changed, but the factual benchmark for meeting that standard is evolving because AI is compressing the time available to identify and remediate vulnerabilities. For Australian boards, the question is no longer whether the organisation has reasonable cyber controls, but also whether those controls can operate at the speed at which AI may discover weakness.
  • This article highlights a series of practical steps to address the evolving threat landscape as well as questions to ask of existing internal policies and approaches.  Leaders who act now will reduce exposure, strengthen resilience, and build confidence with customers, partners, and investors. Those who delay will face growing and avoidable risk.

Cybersecurity capability of AI is increasing rapidly

In June 2026, regulators, frontier AI developers and cybersecurity agencies all reached a similar conclusion: AI is accelerating the discovery of software vulnerabilities faster than many organisations can remediate them. This article provides the context and outlines the regulatory framework.

Specifically, on 7 April 2026, Anthropic had announced its new frontier AI model, Claude Mythos Preview.  Anthropic described Mythos as unusually capable at cybersecurity tasks. In pre-release testing with minimal human steering, the model had identified serious software vulnerabilities across major operating systems and browsers and developed working proof-of-concept exploits.

Recognising the risk if this capability was universally available, Anthropic limited access to around 50 major technology and cybersecurity organisations to enable them to find and remediate vulnerabilities before malicious actors can exploit them. Anthropic also announced funding and support for a cybersecurity coalition focused on these issues, known as ‘Project Glasswing’.

The early results illustrated both the promise and the problem. Anthropic reported that Project Glasswing identified more than 10,000 high or critical severity vulnerabilities during its initial deployment. These figures describe a paradigm shift rather than an incremental change.

During June and July 2026, a series of events occurred:

  • On 2 June, Anthropic announced a major expansion of Project Glasswing, lifting participation from around 50 to approximately 200 organisations across more than 15 countries, adding critical infrastructure operators and government bodies.
  • On 9 June, Anthropic publicly released Claude Fable 5 as a Mythos-class AI with additional safeguards, and provided access to Mythos 5 to a small group of “cyber defenders and infrastructure providers”.
  • On 12 June, the US government acted, citing national security concerns. An export control directive ordered Anthropic to suspend access to both Fable 5 and Mythos 5 for any foreign nationals. As Anthropic could not differentiate between foreign and US users in real time, it closed access to both Fable 5 and Mythos 5 globally.
  • On 15 June, the EU called for global co-operation in frontier AI development.
  • On 22 June, the leaders of the ‘Five Eyes’ cyber security agencies issued a joint statement titled “The AI shift in cyber risk: Why leaders must act now’ (“Joint Statement”).
  • On 26 June, OpenAI previewed its next-generation GPT-5.6 models, led by its flagship ‘Sol’ model. At the request of the US Government, initial access was limited to a small group of vetted partners, the first time a frontier model had been released under a government-managed access list.
  • On 30 June, the US Government withdrew the 12 June export controls.  Following commitments by Anthropic to proactively detect and address security risks, to work with the US Government on protocols, standards and releases, and to report malicious activity, a licence is no longer required to export, reexport or transfer the Mythos or Fable models. The US Government reserved the right to reimpose controls should circumstances change.
  • On 1 July, Anthropic restored Fable 5 globally and restored Mythos 5 access for select organisations.  Anthropic subsequently published details of Fable 5’s cyber safeguards and an early draft of an AI jailbreak severity framework developed with its Glasswing partners, together with a program through which security researchers can report potential cyber jailbreaks.
  • On 9 July, OpenAI made GPT-5.6 publicly available, following further testing conducted by the US Department of Commerce.

These interventions mark a new dynamic. Access to the most capable models is now shaped by government decisions taken on national security grounds, and not by commercial considerations alone. For boards, this is a further source of uncertainty as to when, and through whom, the organisation will gain access to the tools its adversaries may also be seeking.

Urgent 'call to action' to strengthen cybersecurity

Given this context, the Joint Statement is unusually blunt in its wording and urgent in its tone, clearly illustrating the depth of the concern. The Joint Statement sets out a ‘call to action’, states that ‘the urgency is clear’, states that ‘a whole-of-organisation and whole-of-society response is required’, and sets out key actions for corporate leaders. The Joint Statement concludes that ‘we must act now’.

Relevantly, the Joint Statement is expressly directed at boards and corporate executives and comments:

“Cyber risk can no longer be treated as a purely technical issue. This is a core business risk and leadership responsibility. Boards and executives should ensure cyber resilience is in place and works under pressure. It is not enough to have controls. Leaders must be confident those controls will perform during a real incident. This requires reassessing long-standing trade-offs and using  AI  deliberately to strengthen defence – not just improve efficiency”

For Australian boards, the central question is no longer whether the organisation has reasonable cyber controls, but whether those controls can operate at the speed at which AI discovers weaknesses. This is a critical takeaway – the legal standard has not moved, but the benchmark for meeting it has. The benchmark for effective governance is now set against a threat environment moving materially faster than the one that most current frameworks were built for.

In guidance issued on 27 May 2026, the Australian Signals Directorate (ASD) noted that while AI can help identify cyber vulnerabilities, it does not dissolve organisational bottlenecks in remediation.  In essence, it is now far easier to discover problems, than to fix them. As we understand it, only a small share of vulnerabilities surfaced in Glasswing’s initial months have been patched to date.

Two consequences follow as advanced cybersecurity capability becomes a feature of mainstream AI:

  • First, coordinated programs are now using AI to systematically hunt for security flaws within critical software. Organisations can expect a steady stream of newly announced vulnerabilities to enter the public domain. Risk exposure may be concentrated in legacy systems, bespoke applications, and suppliers sitting outside any AI-assisted defensive coalition.
  • Second, AI use by malicious third parties will likely reduce the interval between a vulnerability existing and being actively exploited from weeks to hours. Many governance frameworks remain calibrated to monthly risk reporting, quarterly board reporting and annual cyber reviews: this cadence may no longer be adequate.

The change is, fundamentally, an economic one. For decades, finding a serious vulnerability took scarce expertise and time, which gave defenders some breathing room.  Frontier AI erodes that advantage by lowering the cost of discovery for defenders and attackers alike. Remediation has not become any faster, and it remains bound by change approvals, testing, maintenance windows and, often, customer notice. The scarce resource is no longer the ability to find weaknesses, but the capacity to act on what is found.

Bearing these issues in mind, the Joint Statement sets out five practical actions for organisations, as follows:

  1. Reduce your attack surface: Limit unnecessary system access and external connectivity. Challenge whether systems need to be exposed at all and isolate those that do not.
  2. Accelerate patching processes: AI is shortening the time between vulnerability discovery and exploitation. Delays in patching increase risk, especially for operational systems with long update cycles. Prioritise security updates accordingly to manage risks.
  3. Address legacy systems: Unsupported systems are easy targets. They are not just technical debt, they are strategic liabilities.
  4. Review and strengthen identity and access controls: Limit who can access critical systems. Enforce strong authentication and regularly review permissions.
  5. Prepare for incidents before they happen: Test response plans, train and prepare teams, and assume breaches will occur. Focus on fast containment and recovery.

APRA and ASIC have also highlighted the risk

Pre-dating the Joint Statement, Australian regulators had already moved quickly to respond to the emergence of frontier AI.

On 30 April 2026, the Australian Prudential Regulation Authority (APRA) wrote to all regulated entities with findings from a targeted supervisory review of large banks, insurers and superannuation trustees. APRA’s conclusion was that AI governance, risk management, assurance and operational resilience are not keeping pace with adoption. APRA stated that it expects Boards to have sufficient understanding of AI to be able to set strategy, and provide effective challenge and oversight of AI-related risks.

APRA confirmed that its existing prudential standards already apply to AI – in particular, CPS 234 (information security), CPS 230 (operational risk management) and CPS 220 (risk management) – and must now be operationalised accordingly. The widest gap APRA identified was in third party and supply chain risk: for many entities, the most material AI exposure now sits in vendor platforms with opaque upstream dependencies.

Shortly afterwards, on 8 May 2026, the Australian Securities and Investments Commission (ASIC) issued an open letter to all licensees and market participants warning that misuse of frontier models could expose vulnerabilities at unprecedented speed, scale and sophistication, and urging entities not to wait for advanced tools before strengthening cyber resilience fundamentals.

ASIC commented that it treats cyber resilience as a core licensing obligation, not an IT issue, and asked that the letter be tabled at board and risk committee level. Its assessment was unusually direct: “the clock is at a minute to midnight”.

APRA returned to the theme on 5 June, telling the Senate Economics Legislation Committee that AI adoption is accelerating, governance is not keeping pace, and it is sharpening its supervisory focus on AI governance and supplier risks.

On 1 July, the transition deadline for pre-existing arrangements with material service providers under APRA’s Prudential Standard CPS 230 ended. Regulated entities that have not yet mapped their AI dependencies and fallback options against these requirements should remediate as a priority.

Underneath these regulatory initiatives, the legislative framework continues to strengthen:

  • The Security of Critical Infrastructure Act 2018 (Cth) (SOCI Act) now reaches assets across 11 sectors, including data storage and processing, and brought telecommunications carriers within its risk management regime from April 2025.
  • Australian privacy reform continues in tranches: the Privacy and Other Legislation Amendment Act 2024 (Cth) introduced expanded enforcement powers for the Information Commissioner, a statutory tort for serious invasions of privacy, and criminalised doxxing. Transparency obligations for automated decision-making apply from December 2026.
  • On 2 December 2025, the National AI Plan confirmed the Government will rely on existing, technology-neutral laws supplemented by guidance and an AI Safety Institute, rather than a dedicated AI Act.

The regulatory trajectory is clear. Prudential regulation, privacy regulation, critical infrastructure regulation and corporate governance are converging to treat cyber resilience, AI governance and operational resilience as facets of the same problem. The practical steps and benchmarks required of organisations to address cyber risk are evolving to keep pace with the increasing threat.

The Joint Statement expressly indicates that “cyber risk can no longer be treated as purely a technical issue”. The Joint Statement instead indicates that it is a leadership responsibility for boards and executives. However, the risk is not just one of appropriate leadership, but also involves compliance with legal obligations.

Recent case law and enforcement activity highlight the practical application of the legal obligations. The central legal standard for directors, being the duty of care and diligence, has not changed. What has changed is the benchmark against which that standard is assessed and hence the practical steps required to meet that benchmark.

In ASIC v FIIG Securities Limited [2026] FCA 92, the Federal Court ordered an agreed penalty of $2.5 million on an AFS licensee for cyber security failures, the first such penalty under the general licensee obligations in section 912A, together with a court-supervised programme requiring the engagement of an independent cyber security expert. Importantly, the Court accepted that a successful cyberattack does not, of itself, establish a contravention, noting that:

“It might be observed, at this point, that the mere fact of a successful cyberattack on an entity’s information technology systems does not necessarily indicate that the entity had failed to meet the statutory obligations imposed upon it concerning the protection of its information.”

The question is instead whether the licensee’s measures were adequate to the nature, scale and risk profile of its business, and whether their effectiveness can be demonstrated.  That emphasis on demonstrable adequacy is the same benchmark that AI-accelerated threats are now raising. ASIC’s third cyber action, against Fortnum Private Wealth (filed in mid-2025 and ongoing), is focused not on technical settings but on the adequacy of the licensee’s policies, resourcing and oversight of its authorised representatives. The progression from RI Advice in 2022 to FIIG and Fortnum confirms that cyber resilience is treated as a core licensing obligation enforceable by civil penalty, and that governance and oversight failures, and not only technical ones, will attract regulatory action.

On 5 March, the Federal Court in ASIC v Bekier [2026] FCA 196 applied the duty of care under section 180(1) of the Corporations Act 2001 (Cth) to a series of governance failures (see our recent article on the case here). The Court found in first instance that the CEO, and the general counsel and company secretary had each breached that duty by failing to recognise and escalate serious, foreseeable risks and failing to ensure the board was properly informed. On 17 June 2026, the Court imposed a pecuniary penalty of $700,000 on the CEO (Mr Bekier) and disqualified him from acting as a director of a company for a period of 6 years. Ms Martin (Company Secretary and General Counsel) received a penalty of $400,000 and was disqualified for a period of 7 years. They were also ordered to pay nearly half of ASIC’s costs.

The penalties imposed send a strong message to executives to take reasonable steps to ensure the board is comprehensively informed of the risks and obligations concerning the organisation. As such, executives must escalate red flags and a board may rely on management doing so, but only until obvious warning signs demand harder questions.

Although the case did not concern cyber security, it illustrates the lens through which a future cyber governance failure is likely to be assessed. If exploitation can follow discovery within hours, the question of whether the board received appropriate information, funded cyber protection adequately, and acted with appropriate speed is no longer answered by a quarterly cyber update. Management needs to keep the board appropriately informed. Consequently, boards are under greater pressure to ensure they are actively engaged in the information presented to them, and to interrogate, probe, and where necessary, challenge management, by asking difficult questions about business operations (ASIC v Bekier at [1945], [1951]).

The deeper point reaches beyond directors’ duties.  Australian cyber regulation is largely principles-based: whether the question arises under Australian Privacy Principle 11, CPS 234, the Corporations Act, a contractual warranty, or SOCI, the recurring test is whether reasonable steps were taken. What counts as reasonable has always been judged against prevailing threats, available technology and accepted industry practice.

If frontier AI materially alters the speed and scale at which vulnerabilities are discovered, it shifts the factual benchmark against which reasonableness is measured. Controls which have previously been accepted as adequate may not be regarded as adequate once materially faster detection and remediation capabilities are widely available. That is the mechanism by which a settled legal standard tightens without a word of the statute changing.

The operational questions are pointed. If an incident unfolds over hours rather than days, who has authority to take systems offline? Who notifies third parties, instructs external counsel, contacts insurers and escalates to regulators before the next board meeting? If an emergency meeting would first need to be convened, the decision-making framework may need work.

This points to the critical issue of decision rights. Many organisations have the technical capability to contain a serious incident. The harder problem is organisational, and the people authorised to act are frequently not those who first detect the threat. As response windows shrink, governance will be tested less by technical capability than by the speed at which authority can be exercised. The practical remedy is mundane but frequently absent: standing, pre-authorised delegations to contain, disclose and escalate, settled before an incident and rehearsed ahead of time.

The Joint Statement highlights this issue in the following terms: “Success will not come from having the most tools. It will come from getting the basics right, acting quickly, and integrating cyber security into core business strategy” and “cyber resilience is not an IT issue - it is central to operational continuity and market trust.”

Practical guidance for boards and corporate counsel

The legal standard is settled. However, the benchmark for meeting that legal standard has shifted and hence the practical steps required to meet that benchmark may need to be reassessed.

The Joint Statement sets out a call to action and stresses that the evolving landscape of AI is rapidly changing cyber risk. The Joint Statement sets out three core principles:

  • Secure-by-design and secure-by-default must become standard practice – not an aspiration;
  • Resilience cannot depend on a single solution or technology. Defence in depth remains essential; and
  • As AI systems evolve, new and previously unknown vulnerabilities will emerge, including zero day vulnerabilities.  Breaches will occur. Preparedness helps you contain them quickly and prevent escalation into major operational and financial crises.

As can be seen from the overview set out above, there is a significant amount for boards and corporate counsel to consider. Moreover, the threat landscape is continuing to evolve in unpredictable ways that is causing alarm for the agencies most closely involved in cyber security and national resilience. So what practical steps should boards and corporate counsel now take?

We recommend that boards and corporate counsel test their existing cyber risk management frameworks against the following key questions:

  • Risk tolerance: Has the board reassessed its cyber risk appetite and operational resilience assumptions in light of AI-accelerated threats?
  • Information asymmetry: Are we inside any early-warning or coordinated defensive program, directly or through our key vendors, or do we have to wait for public advisories? Recent events confirm that government decisions may also determine who receives access to frontier models and when.
  • Management reporting: Are directors receiving cyber risk information at a frequency and level of detail appropriate for the current threat environment?
  • Remediation throughput: How many critical vulnerabilities can we verify, patch and validate per week, and is discovery now outpacing that rate?
  • Disclosure readiness: If a systemic flaw is identified in our environment, who decides whether the market, regulators, customers or counterparties must be told, and how quickly?
  • Incident authority: Who is pre-authorised to contain an incident or take critical systems offline before the next board meeting, and is that authority documented and tested through simulation?
  • Supplier exposure: Which material suppliers sit outside any coordinated program, and what assurance exists over them? For APRA-regulated entities, are key AI providers mapped against the CPS 230 material service provider requirements following the 1 July 2026 transition deadline?
  • Board challenge: Can the organisation provide evidence that the board interrogated management’s cyber assurances?
  • Insurance: Are cyber, directors and officers, professional indemnity and technology errors and omissions policies still fit for AI-accelerated events, including aggregation and silent-cyber assumptions?
  • Transactions: Does cyber due diligence now probe patch latency, remediation throughput and governance maturity, not only breach history?

Concluding observations

Frontier AI does not change the law.  However, it does change the factual assumptions on which existing legal standards operate. The Joint Statement and other regulatory responses illustrate the level of concern and the urgency. Even in the weeks covered by this article, the position has continued to shift, with governments intervening directly in how the most capable models are released. The only safe working assumption is that the benchmark will keep rising.

The challenge facing boards and corporate counsel is less technological than organisational. The organisations best positioned for this next phase will not necessarily be those with the most sophisticated technology. They will be those that can recognise, escalate and remediate vulnerabilities within timeframes regarded as reasonable in the evolving threat environment.

The defining regulatory question for boards is no longer whether vulnerabilities exist: they always have. The question is whether the organisation can learn of them, assess them and respond before an attacker does. Frontier AI is compressing the timeframe to answer that question.  As the Joint Statement notes: “Leaders who act now will reduce exposure, strengthen resilience, and build confidence with customers, partners, and investors. Those who delay will face growing and avoidable risk.”

Contact