By James Dickson, Founder and Managing Director, Oceanic Consulting Group
Frontier AI can now identify day-zero software bugs and backdoors no software tester could conceive of that have lain dormant for decades. Market participants should look through their technology stack, and behind their vendors, before a regulator asks them to.
In October I sat on a panel in Kuala Lumpur on the future of market structure. The questions were about T+0, 24/7 trading and multi-asset settlement. My answer was that the infrastructure is not ready. In every market, including ours, a large part of the post-trade stack runs on code and hardware that is 20 to 30 years old. Books and records systems should be big, stable and boring. Many of them are also out of vendor support, and in the last six months that has become a cyber problem as well as a budget problem.
Project Glasswing, April 2026
On 7 April 2026, Anthropic announced Project Glasswing and gave eleven launch partners, mostly large technology companies plus one bank, and more than 40 other organisations, access to Claude Mythos Preview. The model found thousands of high-severity vulnerabilities, including some in every major operating system and browser. A 27-year-old bug in OpenBSD, an operating system used in firewalls and network equipment, let an attacker crash any machine running it by connecting to it. In the Linux kernel, which underpins the trading platforms at exchanges such as the NYSE, Nasdaq and the London Stock Exchange, the model combined several vulnerabilities to take full control of a machine from an ordinary user account. Jamie Dimon, CEO of JPMorgan, the bank in the group, said on his first-quarter earnings call that Mythos ‘does create additional vulnerabilities’. Attackers will have the same capability, if they do not already.
APRA wrote to every regulated entity on 30 April calling for a step change in how they manage AI-related risk, naming frontier models such as Mythos. ASIC wrote to all AFS licensees and market participants on 8 May. ASIC Commissioner Simone Constant said the clock is at a minute to midnight.
PSPF Direction 002-2026
On 29 September, the Secretary of Home Affairs issued Direction 002-2026 under the Protective Security Policy Framework, which binds non-corporate Commonwealth entities, such as departments and most agencies. It says the Commonwealth’s technology estate has been targeted using frontier AI capabilities and that running vulnerable legacy systems is an unacceptable risk. Every entity must stocktake its legacy technology, including systems run on its behalf, prepare a Legacy Technology Risk Management Plan with a reduction target, and report back to Home Affairs by 31 March 2027. It binds those agencies only, but I read it as a preview of what our regulators will ask for next.
Figure 1: Legacy technology and cyber exposure, April 2026 to March 2027. Source: Oceanic Consulting Group.
Obligations already in force
The technological and operational resilience chapters of the ASIC Market Integrity Rules have applied since March 2023. They require participants to identify their critical business services, assess the information and IT assets that support them, carry out due diligence on the third parties that provide or operate them, test business continuity plans at least annually, and notify ASIC immediately of a major event, including a cyber attack. A stocktake of the estate, including what vendors run for them, is the first step in meeting those rules; the register and the board paper behind it are the evidence to keep.
On 24 September, ASIC announced amendments that modernise the trading system obligations for algorithmic and AI-enabled trading, commencing on 18 March 2028. A participant must test a trading algorithm before first using it or providing it to a client, and again before any material change. ASIC’s draft update to Regulatory Guide 265 addresses vendor-supplied algorithms and systems directly, and feedback is due by 5 November. Eighteen months is not long if the system to be tested sits on a platform nobody can patch.
For APRA-regulated groups, CPS 234 has required since 2019 that information assets, including those managed by third parties, be classified by criticality and sensitivity. CPS 230 has required since July 2025 that critical operations, and the material service providers that support them, be managed.
In February, the Federal Court ordered FIIG Securities to pay a $2.5 million penalty. The failures FIIG admitted included operating systems not on vendor-supported versions and patches for vulnerabilities disclosed in 2017 and 2019 left unapplied for about four years.
Look behind the vendor
The gap we see most often is the system a firm does not know it has, because it sits inside a vendor product: a modern front end on an order management core from the 1990s. APRA’s April letter, writing about AI adoption, described a related habit as ’an overreliance on vendor presentations and summaries without sufficient examination of key AI risks’.
When a vendor says a platform is supported, we recommend clients ask what that means for the components underneath it: which operating system, database and middleware versions, who patches them and how quickly after a critical security advisory, and whether the vendor could apply that patch without breaking something. If the vendor cannot answer, use the audit rights in the contract and raise it at renewal.
Vendor roadmaps and T+1
Vendor selections being made now are, in some cases, for a post-2030 shift, and once live you will be on that platform for a decade, or more. You should be asking every technology partner whether they are on the bleeding edge or behind it: whether the move is onto a modern stack, and whether they have planned for T+1 or are waiting until after CHESS Replacement Release 2 in 2029. ASX’s position is that T+1 follows Release 2 by a minimum of 12–18 months. If nobody has run the code through frontier AI tools and the platform plan stops at Release 2, have you fully considered the decision you are making and the future disruption to your business?
What to do now
- Stocktake the whole estate, including everything run for you by vendors, outsourcers and cloud providers, and record vendor support status and end-of-life dates against every system.
- Decide what is legacy: anything out of vendor support that is also impractical, uneconomic or above risk appetite to keep.
- Set a reduction target and a date, approved by the board against its risk appetite.
- Prioritise by exposure: internet-facing systems, critical business services, custody and managed account platforms, and anything holding client or identity data.
- Mitigate what has to stay: isolate it on the network, keep it off the internet, restrict privileged access, and give it a retirement date and an owner.
- Keep the evidence in a form you could hand to ASIC tomorrow.
The Commonwealth has been told in writing that running vulnerable legacy systems is unacceptable, and given a date. Nobody has written to market participants in those words yet, but the Market Integrity Rules, ASIC’s May letter and the September amendments all point the same way. When the request comes, the firms that have done the stocktake will send it in and get on with their day. The rest will be caught flat-footed.
James Dickson is Founder and Managing Director of Oceanic Consulting Group, a practitioner-led financial services consulting firm operating across Australia, the UK, Singapore, Hong Kong and Dubai.
DOWNLOAD SIAA Monthly – November edition PDF
BACK TO SIAA MONTHLY – NOVEMBER 2026