AI Next WaveAugust 31, 2026
AI Can Speed Up the Attack. One Supplier Can Still Stop the Work.
The chair of a global financial watchdog says advanced AI could magnify cyber disruption across shared providers. OpenAI’s planned cutoff of Cursor gives companies a live example of dependency risk, while Europe and NASA are building different models for shared technology capacity.
The Short Version
The next AI failure may spread through companies that never chose the same model.
Andrew Bailey, chair of the Financial Stability Board, warned G20 finance ministers Monday that frontier AI could change the speed, scale and economics of cyber risk. Financial firms share cloud platforms, software suppliers, data networks and other technology providers. A faster attack or rushed patch can therefore move across institutions and borders.
Bailey called AI-enabled cyber risk the financial system’s most immediate concern from frontier models. He also singled out concentration among third-party technology providers and urged firms to prepare for simultaneous disruption, including the ability to rebuild critical systems and data from clean infrastructure.
The warning is forward-looking. It provides no measured increase in AI-caused financial losses and creates no new binding rule. It still identifies a practical weakness that reaches well beyond banking. An organization can spread work across many AI applications while those applications depend on the same model company, cloud or identity service.
OpenAI supplied a vivid example Friday. The company said it intends to stop providing models to Cursor after SpaceX acquired the coding-tool company. The proposed shutoff date is November 12. Cursor offers models from several providers, and Anthropic says it will add capacity for Claude. The product can continue. Workflows built around OpenAI models may still need to be retested.
Europe is taking a public-infrastructure approach. EuroHPC signed a €387.8 million contract for LUMI-AI, a supercomputer intended for researchers, startups and industry. It is scheduled for the second half of 2027, so access and economic results remain ahead.
NASA’s Roman Space Telescope offers another model. The observatory launched successfully Sunday and will produce far more data than most researchers can download to their own computers. NASA’s partners are pairing public data with cloud-based tools that let people bring analysis to the archive.
Capability gets the attention. Continuity depends on who controls the service, how work moves when a provider changes course and whether people can reach the data and tools needed to recover.
Finance Prepares for Faster Disruption
The FSB puts shared providers inside the risk scenario
The Financial Stability Board coordinates international work on risks to the global financial system. Its chair’s letter ahead of the August 31 and September 1 G20 meetings places frontier AI inside that mandate.
Bailey describes two pressures arriving together. Advanced models may help attackers find vulnerabilities and operate at greater speed and scale. Defenders may also use the same capabilities to inspect systems and patch flaws faster. The resulting tempo can overwhelm change controls, testing and recovery processes.
Shared infrastructure makes the problem systemic. Banks, insurers, payment companies and market operators often depend on a relatively small group of cloud, software and network providers. A disruption at one common supplier can affect firms that otherwise appear separate. Differences in national cyber capability and recovery rules can carry the incident across borders.
The letter asks financial institutions, market infrastructures and technology providers to strengthen vulnerability management, response and recovery. It specifically raises the ability to restore critical systems and data from bare metal. In plain English, a firm should be able to rebuild from known-clean hardware and software when trust in the existing environment has collapsed.
That standard exposes the weakness in many resilience plans. A backup stored inside the same cloud account, accessed through the same identity service and restored with the same compromised automation may preserve the dependency that caused the outage.
Banks and regulators should test a combined scenario: an AI-assisted attack finds flaws quickly while a shared technology provider is unavailable or untrusted. The exercise should cover customer access, payments, data reconciliation, manual service, public communication and restoration from clean systems. Useful measures include the time to isolate the problem, the time to restore priority services, transactions that need repair and records that cannot be reconstructed.
Small financial firms face a sharper tradeoff. Shared platforms give a credit union, local bank or fintech capabilities it could never build alone. The same arrangement leaves fewer technical alternatives during an incident. A realistic plan identifies which services can run manually, which provider has to recover first and how customers receive accurate information when the main channel is down.
Bailey’s letter is a supervisory warning, not proof that frontier AI has already increased financial losses by a particular amount. The FSB says it is exploring safe deployment of frontier models for cyber defense and ways to improve recovery. Institutions still need incident data, exercises and independent testing to turn the warning into operating evidence.
Cursor Gets a Deadline
A model supplier can change a product without changing its code
OpenAI said Friday that it intends to wind down its contract supplying models to Cursor, with a proposed shutoff date of November 12. The announcement followed SpaceX’s acquisition of Anysphere, the company behind Cursor.
OpenAI said its contract allows cancellation for a limited period after a change of control. It cited concerns that SpaceX would use its technology outside OpenAI’s terms and said future OpenAI models would be withheld from Cursor.
Those are OpenAI’s stated reasons. SpaceX chief Elon Musk disputed the company’s account, and Cursor co-founder Michael Truell said the companies were talking about a resolution. The cutoff is planned rather than completed.
Cursor already supports models from Anthropic, Google, SpaceXAI and OpenAI, along with its own models. Anthropic said it would increase computing support for Claude inside Cursor. Users therefore have alternatives, and the announcement provides no evidence that Cursor itself will stop working.
Substitution still carries work. Coding models can differ in the files they change, the tests they run, how often they ask for approval, how they handle long projects and the kinds of errors they produce. A team that has tuned instructions, review rules and cost expectations around one model may see quality, latency and spending change after a switch.
The practical unit of dependency is the accepted result. A software team should keep a small evaluation set drawn from its own work: a bug fix, a test-writing task, a security review, a modest refactor and a documentation change. Run the same tasks through the primary and fallback models, then compare accepted output, reviewer time, defects, tool calls, latency and full cost.
Portability also requires records. Teams need exports of prompts, rules, evaluation cases and activity logs in formats they control. They should know whether a model can be changed through configuration or whether its behavior is embedded in application logic, billing, permissions and employee habits.
The Cursor episode carries a broader message for businesses buying AI through another product. The interface may come from one company while the core capability comes from several upstream providers. Procurement should identify each critical supplier, the notice period for removal, the data that can leave with the customer and the performance that has to be revalidated after a change.
Europe Buys Public AI Capacity
LUMI-AI is a signed contract with delivery still a year away
EuroHPC has signed a €387.8 million contract with French state-owned Bull to build LUMI-AI in Kajaani, Finland. The price includes acquisition, delivery, installation and maintenance. Funding is split evenly between EuroHPC and a six-country consortium covering Finland, Czechia, Denmark, Estonia, Norway and Poland.
The system is scheduled for deployment in the second half of 2027. Bull and the LUMI consortium say it should provide ten times the AI capacity and nearly twice the conventional high-performance computing capability of the current LUMI supercomputer. Those are project specifications and vendor expectations, since the machine has yet to be built and accepted.
LUMI-AI will use AMD processors and accelerators, IBM storage and Nokia networking. Its direct liquid cooling is designed to feed excess heat into Kajaani’s district-heating system, and the operators say the data center will use renewable electricity.
The hardware choice broadens Europe’s supply base beyond the dominant Nvidia stack. The ownership model also matters. EuroHPC is building shared computing infrastructure for academic research and industrial AI, with public partners carrying the capital cost.
Shared capacity can help a startup, university or midsize manufacturer run work that exceeds a local cluster or a normal cloud budget. LUMI’s operators say the AI Factory will support startups and small and midsize companies. The contract announcement does not yet provide the allocation rules, queue times, support levels or costs that will determine who receives useful access.
The public return should be measured after deployment through accepted projects, organizations served, waiting time, repeat users, research completed, products launched and skills retained in the participating countries. Processor capacity and contract value describe the input.
Applicants can prepare before the machine arrives. A credible project needs a defined workload, a lawful and usable dataset, a baseline from smaller infrastructure and a reason that more computing capacity changes the outcome. Teams should also know what they will do if access arrives later than expected or if the workload needs a different software stack.
A Space Telescope Brings the Compute to the Data
Roman’s public archive was designed into the mission
NASA’s Nancy Grace Roman Space Telescope launched Sunday at 7:26 a.m. Eastern time aboard a SpaceX Falcon Heavy. The observatory separated from the rocket, established communications and deployed its solar panels and lower sunshade.
Roman is now on a three-month journey and commissioning period toward the second Sun-Earth Lagrange point, about one million miles from Earth. Its instruments still need calibration and testing. NASA expects the first images in early 2027.
The roughly $4 billion observatory combines infrared sharpness comparable to Hubble with a field of view that can survey the sky far faster. Its five-year primary mission will study dark energy, dark matter, exoplanets and many other objects.
The data system may prove almost as consequential as the telescope. NASA says Roman will transmit about 1.4 terabytes each day. The Space Telescope Science Institute expects about 20 petabytes during the first five years. All Roman science data will be immediately public, without an exclusive period for the team that collected it.
Downloading that archive to a typical university or personal computer would be impractical. The Roman Research Nexus lets scientists run code near the cloud-hosted data. It is already available with simulated datasets, tutorials, analysis tools and editable algorithms.
That design widens participation while creating a new skill requirement. Astronomers will need facility with cloud analysis, reproducible software and large-scale data quality. Software engineers, statisticians and machine-learning specialists can contribute to discovery pipelines that surface unusual events for expert review. NASA expects AI, machine learning and citizen scientists to help sift the stream.
Equal access also requires training, computing allocations, reliable tools and time to interpret results. The Nexus can reduce the need for local hardware. Its real value will appear in who can use the archive, how reproducible the analyses are and whether discoveries reach beyond institutions with the largest computing teams.
Roman also provides a useful pattern for public technology projects. The instrument, archive, software and user access were designed as one system. A costly capability produces broader value when people can reach its output and work with it.
Opportunity Radar
AI continuity drills for organizations without a large technology staff
Small banks, insurers, healthcare providers, professional firms and software companies are adding AI through products that depend on several upstream model, cloud and identity providers. Many have an incident plan for a server outage and no tested plan for a model removal, policy cutoff or simultaneous provider failure.
A cybersecurity consultancy, managed service provider or AI implementation specialist could run a focused continuity drill around one important workflow. The engagement would map upstream dependencies, export the organization’s rules and evaluation cases, switch to a fallback model, restore access from clean credentials and document the decisions needed during the change.
The buyer gains evidence about whether the workflow can continue and what quality or cost changes under the fallback. The provider must validate that the exercise tests real work without exposing production data or promising regulatory compliance beyond its qualifications. Time to switch, accepted output, reviewer effort, unresolved errors and restored records offer a useful scorecard.
What You Can Do With This
If AI sits inside a critical workflow
Name the model, cloud, identity service, data source and application owner behind it. Record the notice and export terms for each supplier. Run a real task through a fallback and compare the accepted result, review burden, cost and speed.
If you manage cyber resilience
Add a common-provider outage to the next exercise. Assume a faster vulnerability cycle and test isolation, manual service, customer communication, data reconciliation and restoration from clean infrastructure.
If you hope to use public computing capacity
Prepare the work before the machine. Build a smaller baseline, document the dataset, estimate computing needs and define the outcome that added capacity should improve. Track LUMI-AI access rules as they emerge.
If you work with public data
Try Roman’s simulated datasets and cloud tools before the observatory begins science operations. Practice reproducible analysis near the archive and preserve the assumptions, code and versions behind each result.
The Bigger Picture
The financial system’s AI exposure reaches beyond malicious prompts and model errors. Shared technology providers can transmit a disruption across firms, and supplier decisions can remove a capability from a product even when its application code remains intact.
The FSB’s warning gives institutions a harder recovery standard. Cursor’s deadline shows why model portability needs to be tested before a cutoff. Europe’s LUMI-AI contract creates public capacity, with access and results still to prove. Roman pairs an expensive instrument with immediate public data and cloud tools built for the scale of its output.
Each case turns capacity into a chain of dependencies. The chain includes contracts, credentials, data rights, computing infrastructure, human review and a route back to service.
Organizations can make that chain visible now. A useful map names the provider, the work it enables, the fallback, the evidence that the fallback works and the person authorized to make the switch.
AI can move work and risk faster. Recovery still depends on preparation done at human speed, before the pressure arrives.
References
Cursor: Current supported model providers and pricing structure
Reuters: Europe expands its public AI computing network with LUMI-AI, August 31, 2026
NASA: Roman Space Telescope launch, commissioning plan and expected data rate, August 30, 2026
Reuters: Roman’s launch, cost, mission length and repurposed satellite origin, August 30, 2026
Space Telescope Science Institute: Roman Research Nexus and five-year data scale, April 16, 2026
Roman User Documentation: Immediate public access and cloud-based analysis of Roman data