The Architectural Answer to the Mythos AI Security Vulnerability: GRIDS
Press Briefing for Mainstream Press
QPQ AG, Switzerland – 7 May 2026
Every link in this briefing leads to a primary source. QPQ is contactable for verification of any claim not covered by an embedded link.
The story
On 7 April 2026, Anthropic announced an AI model called Mythos that can break into other people’s computer systems on its own, at machine speed. In Anthropic’s own words, the model can perform “account login bypasses that allow unauthenticated users to log in without knowledge of their password or two-factor authentication code.” In plain English: it logs into other people’s accounts without knowing the password, and the code the bank texts to a phone does not stop it. To prove the point, the model identified a 27-year-old flaw in OpenBSD – one of the most secure operating systems on the internet – and used it to take complete control. Anthropic’s own estimate is that within six to eighteen months, equivalent capability will be in the hands of others who will not behave responsibly.
Within days, the US Treasury Secretary and the Chair of the Federal Reserve convened Wall Street’s largest bank chief executives in the first joint emergency meeting of its kind since the financial crisis of October 2008. The Bank of Canada convened its Financial Sector Resiliency Group. The Bank of England is convening its Cross Market Operational Resilience Group.
On 13th April the Cloud Security Alliance, SANS, and OWASP jointly published an emergency framework: eleven priority actions, with the report’s own caveat that “long-term goals should be considered a quarter away at most.” Two days later, the UK government’s open letter to business leaders recorded the UK AI Security Institute’s assessment that frontier AI capabilities are now doubling every four months, against the previous estimate of every eight.
Subsequently, on 24 April, the Swiss Financial Market Supervisory Authority told Bloomberg that “the uncontrolled and immediate availability of AI models such as Mythos would be classified as a systemic risk” because “virtually all existing software systems could simultaneously be affected by a multitude of previously unknown zero-day vulnerabilities, which would be exploited immediately and via AI.” FINMA confirmed it is in contact with banks and “critical service providers” on the matter. On 4 May, the Eurogroup convened in Brussels to discuss Mythos access for European institutions; the Bundesbank President said all relevant institutions should have access to the technology to avoid competitive distortions. Switzerland is not in the Eurogroup process. The audiences at risk are not.
Everything your audience uses on the web today is exposed to a class of attack that the existing defences cannot stop. Their bank account, their pension provider, their tax record, their medical record, their email, their employer’s payroll, the supply chain that puts food on their table – every one of these sits behind credentials that Mythos has demonstrated it can defeat at machine speed and without human oversight. Anthropic has held the model back; on Anthropic’s own timeline, others will not.
The window is six to eighteen months and counting, and almost none of it has been put in front of the people who will live with the consequences.
Why this matters to your audience
Every breach the public has heard about over the last twenty years – banks, retailers, hospitals, government agencies – has had the same root cause: the internet was built to share information, and the thing that proves a customer is who they say they are has had to sit on a computer connected to the internet, because there has been nowhere else to put it. Every defensive measure of the last thirty years has tried to make reaching that proof harder. None has changed the fact that the proof is exposed or can be exposed, and Mythos has now made finding it autonomous.
The cybersecurity industry’s response is to recommend spending more on the same kinds of products it has been selling for thirty years. Global security spending reached $213 billion in 2025 and is forecast at $240 billion in 2026 – 12.5% growth in a single year against a threat that has just rendered the underlying assumption obsolete. The 13 April emergency framework that supervisors and regulators will be citing in the months ahead lists its authors and reviewers on its title page: most are CISOs, vendors, investors in security firms, training organisations, and conference operators whose commercial position is served by an answer that is more of what they sell. Lead author Gadi Evron is chief executive of Knostic, whose tools appear among the recommended options in the framework’s first priority action; the framework’s publishing bodies are themselves named in its adoption pathways. The affiliations are disclosed on the title page; the conflict at the points where the affiliations bear on specific recommendations is not flagged.
The architectural answer
A Swiss company, QPQ AG, has been running an alternative architecture since 22 October 2024. The architecture is called the Internet of Economics. The component that addresses Mythos directly is GRIDS – Gajumaru Remote Instruction Dispatch and Serialisation – a free open protocol released under the GPL3 open-source licence at Main Net on 26 April 2026.
The principle is simple. Nothing on the institution’s computer should be able to act as the customer. The signing key sits inside the hardware-backed secure storage of the customer’s own device – their laptop today, their phone from end Q2 2026 – in a sealed part that the operating system, the browser, and other applications cannot read. When the institution needs to verify the customer or authorise an action, it sends the specific instruction to the customer’s device. The device displays the instruction in plain language: “approve transfer of £500 to John Smith”; “approve change of address”. The customer reads it on their own device’s screen and approves. The device produces a one-off cryptographic signature bound to that specific instruction. The institution’s system verifies the signature against a public counterpart on file – useless to anyone else – and acts on it. There is no password to steal, no code to intercept, and nothing left over after the action that an attacker can take and use to keep pretending to be the customer.
A five-minute live demonstration is available at https://youtu.be/WkzNErEg51o – login, transfer action, and QR code login. It works today on a laptop or desktop; the mobile reference application follows in July 2026. The protocol is open source and free; QPQ IaaS AG, the Swiss operating subsidiary in Einsiedeln, provides the engineering integration for institutions that want it built into their existing systems by the team that built the protocol and runs it for the group’s own operations. The first sovereign user is the Liechtenstein Trust Integrity Network, with Telecom Liechtenstein as majority owner, deploying national infrastructure on this architecture in the second half of 2026.
Our commercial position
The GRIDS protocol is open source and free under GPL3; QPQ does not charge for the protocol or for the reference applications, GajuDesk and GajuMobile.
The commercial offer is engineering integration through QPQ IaaS AG, the Swiss operating subsidiary in Einsiedeln: institutions that want GRIDS built into their existing systems by the team that built it engage QPQ IaaS AG on a project basis. We make that point explicitly because the cybersecurity industry’s framework does not.
The story angles for your audience
The thirty-year mistake that has been costing your audience. Every breach the public has heard about has had the same root cause, and the public has been told it was bad luck or clever criminals. It was neither. It was the wrong tool used for the wrong job, for thirty years.
The disclosure problem at the heart of the cybersecurity policy response. The framework that supervisors and regulators across the major democracies will be citing in the months ahead is being authored by parties whose commercial position is served by the answer being more of what they sell. The work to verify the conflict takes a reporter five minutes from sources already in the public domain.
The silence at the customer end. Treasury, the Federal Reserve, FINMA, the Bank of England, the Bank of Canada, and the European Central Bank have all convened on Mythos. The Eurogroup met on it three days ago. None of these institutions has yet told their customers, citizens, or constituents what they now know about the security of their data. The audience has the right to know.
What this is part of. GRIDS is the first commercial tool of the Internet of Economics. Most of it is not yet public. The Gajumaru base layer has further implications for inter-bank settlement and correspondent banking. A reporter who wants to understand where this leads, beyond the immediate Mythos answer, has the conversation available.
What QPQ can offer your newsroom
On-record interview with Greg Chew, CEO. Live, recorded, or written. The thirty-year architectural error, the disclosure problem at the heart of the cybersecurity industry’s response, what your audience should ask their bank, why this matters now.
Live demonstration of GRIDS. Ten to fifteen minutes, remote. Greg Chew or Craig Everett (Chief Product Officer and the engineer who built the protocol) walking through what an attacker cannot do once GRIDS is in place. Hands-on access for your reporter is available through gajumining.com/downloads; the team will get them signed in within minutes, with no password.
Background interview with the engineering team. Ulf Wiger (Chief Technology Officer, formerly chief designer of Ericsson’s AXD 301), Craig Everett (Chief Product Officer and GRIDS architect), Dimitar Ivanov (Chief Development Officer).
qpq.swiss · gajumaru.io · gajumining.com
QPQ AG (Industriestrasse 47, Zug) built the Internet of Economics architecture and holds the intellectual property. QPQ IaaS AG (Allmeindstrasse 17, 8840 Einsiedeln) is the integration counterparty for institutions deploying GRIDS. Gajumaru and GRIDS operational since 22 October 2024. Main Net: 26 April 2026.