The European Central Bank (ECB) has concluded a weeks-long inspection of Euro banks with a warning: preparedness for the AI security risk is not adequate, and more spending on cybersecurity will be required to get up to speed.
ECB members have been asking Euro banks about their preparedness levels for weeks, concluding with a late May meeting in which outgoing Vice President Luis de Guindos declared that they are broadly not ready for the AI security risk posed by new models like Mythos. Guindos said that spending on defensive systems and timely patching needs to be increased to keep pace with frontier models that are able to very quickly detect vulnerabilities.
Euro banks struggle without access to advanced models
ECB VP de Guindos noted that part of the lack of preparedness for the AI security risk is likely due to lack of awareness of the capability; Euro banks by and large have not been given advance access to Mythos for security testing purposes. An unnamed United States bank that has been testing with Mythos for some time now was invited to the final meeting to brief the Euro banks on some of their findings.
The ECB also stresses that this is not just a matter for large Euro banks. All financial institutions, including small lenders, will need to be equally prepared for the emerging AI security risk. The key finding of this series of inspections is that investment is the most important element. Banks must simply budget more for elements such as faster patching, awareness training and more modern defensive elements suited to matching AI speed and capability.
The ECB noted that Euro banks have been aware of these issues as a future possibility for some years now and have been making some progress in addressing them, but warn that at this point the technology will likely catch and pass them if they do not speed things up. The AI security risk has come into very sharp focus with the recent announcement by Anthropic that while the public release of Mythos still does not have a firm date, comparable models will very likely be available within 6 to 12 months.
Frontier AI security risk is too immediate to ignore
The ECB also pointed out that lack of broad access to Mythos Preview and similar models can no longer be used by Euro banks as an excuse for inaction, as there is a real and demonstrated possibility of unauthorized threat actors getting access to them. A recent example of this AI security risk comes from Anthropic itself, which saw some 500,000 lines of Claude Code leaked due to a simple error in a public npm release.
The new AI models are expected to greatly amplify this risk as they will possess the ability to analyze a target’s full set of vulnerabilities within minutes, including some that may not have been previously documented. Anthropic has said that its own testing with Mythos (under its “Project Glasswing”) has uncovered hundreds of previously unknown vulnerabilities in major browsers like Firefox and across numerous OSS projects.
However, independent testing of Mythos has thus far been kept out of Europe aside from a small amount of select government agencies in the United Kingdom. The European Commission and regional finance ministers have been petitioning Anthropic for access, but to no avail as of yet. Access has mostly been restricted to the US and to its government agencies so far, but a number of major US banks have also been looped in. OpenAI has recently announced that it will make its comparably powerful ChatGPT 5.5 model available to dozens of organizations in the EU, but there have yet to be any serious talks about Mythos access in the region.
Ryan McCurdy, VP, Liquibase, provides some added detail about why frontier model access is so crucial to the finance industry given how heavily it will inevitably be targeted: “The important word here is structural. AI is changing cyber risk from a periodic security problem into a continuous operational one. When advanced models can surface flaws faster, banks cannot rely on manual review cycles and patch programs built for human speed. They need stronger control over what can change, clearer approval boundaries around production systems, and faster ways to contain risk when something goes wrong. Otherwise, the organization is just moving into a higher-speed threat environment with a slower control model.”
The “curl” project potentially provides some evidence as to why access to the latest models is crucial for curbing AI security risk. While Project Glasswing found scads of vulnerabilities in Firefox and thousands of other projects, it found just one lone previously undiscovered vulnerability in curl. That is thought to largely be owed to ongoing preparation by the curl team, which had already been using some of the best available AI models to actively explore for (and patch) vulnerabilities in its over 170,000 lines of code. The curl team revealed in a blog post that it had been using tools like AISLE, Zeropath and OpenAI’s Codex Security months in advance of its probing by Mythos, and had located over 200 bugs that it applied fixes for.
However, Steven Swift (Managing Director at Suzu Labs) notes that most other industries and their organizations will need to apply this same level of diligence when these new models emerge sometime in the coming year: “There’s not much here that’s unique to the banking sector. Most of the AI vulnerability risk that banks and their regulators are concerned about applies broadly to most companies. Every industry wants to keep their data secure, avoid incidents, and minimize impact if and when something does go wrong.“
“The impact will be greatest where there are legacy systems in place that aren’t easily patchable, if and when vulnerabilities in those systems are detected,“ Swift added. “And impact will have the widest reach when new vulnerabilities are detected within the chain of dependencies that software relies on, rather than directly with the code in the application itself.”
