AI & Tech News Roundup -- May 25, 2026

Anthropic's Mythos found tens of thousands of unknown vulnerabilities in the world's banking infrastructure. Google replaced its search box with AI agents. And the US government now pre-screens every major frontier AI model before it ships.

Anthropic's Mythos Found Vulnerabilities in the World's Banking Infrastructure. Now What?

The biggest AI story of the past few weeks isn't a product launch -- it's a security crisis most people outside finance are still catching up on.

Anthropic's Mythos model, a specialized cybersecurity AI, discovered tens of thousands of previously unknown software vulnerabilities across the world's financial and critical infrastructure systems -- decades-old bugs hiding in banking code that had never been caught by conventional security tools. CEO Dario Amodei went public with a stark warning: there is a six- to twelve-month window to patch these vulnerabilities before adversarial actors -- potentially including AI systems from state-level actors catching up to Mythos's capabilities -- can exploit them at scale.

The response has been unusual for an AI story. The Financial Stability Board was briefed at the request of Bank of England Governor Andrew Bailey, who chairs the FSB. Euro-area finance ministers discussed it as a financial stability concern. Anthropic has limited Mythos access to a small group of US companies -- Apple, Amazon, JPMorgan Chase, and Palo Alto Networks -- specifically to reduce the risk of the tool being used offensively while the patching window is still open.

Calling it a "cybersecurity moment of danger," Amodei's framing is deliberate: this isn't theoretical AI risk, it's infrastructure risk with a clock attached.

All 5 Major AI Labs Now Give the US Government Pre-Deployment Access

Catalyzed by the Mythos revelations, Microsoft, Google, and xAI this month joined OpenAI and Anthropic in agreeing to give the US Commerce Department's Center for AI Standards and Innovation early access to frontier AI models before public release.

The agreement is voluntary -- for now. The Trump administration is considering formalizing the process through an executive order, which would transform a good-faith commitment into a regulatory requirement. The five-lab alignment is notable: six months ago, voluntary pre-deployment evaluation was a point of contention across the industry. The Mythos situation appears to have shifted the calculus on whether self-governance is sufficient.

For investors and enterprises watching the regulatory trajectory, this is a meaningful data point. The path from voluntary to mandatory AI oversight is accelerating, and companies building on frontier AI infrastructure should expect compliance requirements to become more structured through 2026 and into 2027. The parallel CLARITY Act moving through Congress reflects the same impulse -- policymakers want frameworks in place before the next capability jump, not after.

Google Replaced Its Search Box

At I/O 2026, Google didn't just update search -- it replaced the fundamental premise of it.

The traditional query box is being replaced by an AI-powered interface built on Gemini 3.5 Flash, featuring "information agents" that monitor the web continuously, generative UI that builds custom tools on the fly, and a conversational mode that expands seamlessly for longer interactions. AI Overviews now reaches 2.5 billion monthly users; AI Mode, the conversational interface launched just a year ago, has already crossed 1 billion monthly users.

Google also announced Gemini Omni -- a model capable of generating "any output from any input" across text, images, audio, and video -- and Gemini Spark, the persistent background agent that monitors connected apps and proactively surfaces insights we covered at launch last week. Spark remains in beta with access limited to Google AI Ultra subscribers and trusted testers.

The speed of Google's AI integration into Search is worth noting. The company went from experimental AI features to replacing its core product interface within 12 months of AI Mode launching. For anyone who thought Google would be cautious about disrupting its own search business, I/O 2026 answered that question.

Enterprise AI Agents: The 88% Failure Rate Nobody Is Talking About

New data from Gartner offers a sobering counterweight to the AI agent hype cycle: 88% of enterprise AI agent pilots fail before reaching production. The cause is not model quality -- it's governance and observability gaps. Organizations deploy agents without adequate frameworks for monitoring what they're doing, auditing their decisions, or overriding them when they go wrong.

The headline adoption numbers are still significant: Gartner projects 40% of net-new enterprise applications will include agentic capabilities by end of 2026, up from under 5% in 2025. Industry surveys report 57% of organizations seeing measurable impact in software development and 55% in customer service.

But the 88% failure stat is the one that matters for anyone building with agents right now. The delta between "runs in a demo" and "runs in production" is governance architecture -- logging, access controls, human-in-the-loop checkpoints, and escalation paths. Companies that solve that layer will own the enterprise AI stack in 2027.

OpenAI Named Gartner Leader in Enterprise Coding Agents

On May 22, Gartner named OpenAI a Leader in enterprise coding agents -- the first time the category has appeared as a discrete Gartner evaluation. The recognition comes as OpenAI and Anthropic are both reportedly exploring acquisitions of engineering services and consulting firms to accelerate enterprise deployment, with OpenAI's Deployment Company raising roughly $4 billion and Anthropic's comparable initiative securing approximately $1.5 billion.

The consulting acquisition push is a significant strategic shift. Rather than waiting for enterprise clients to figure out AI integration on their own, both labs appear to be moving toward owning the implementation layer -- the professional services, fine-tuning, and deployment work that historically goes to firms like Accenture or Deloitte. If that trend continues, the competitive dynamics in enterprise software services are going to look very different by 2027.