TLDR IT 2026-05-12

AI is everywhere right now. But for many teams, reality hasn't matched the promise.

What's actually working?

Tines released a new guide that takes a more practical look at AI adoption for security and IT teams - from a framework for evaluating tools beyond the demo, to human-in-the-loop best practices for safe, scalable AI.

If you're thinking about AI beyond experimentation, this is a useful place to start.

US Bank is shifting critical applications to AWS as financial services firms modernize their infrastructure for AI workloads. It shows how AI adoption is increasingly forcing banks to rethink foundational cloud architecture, not just add tools on top.

A new Alliance for Critical Infrastructure is launching to improve how the US prepares for major cybersecurity crises. The group is focused on changing planning and response models for sectors where outages can have national-level consequences.

Google says it disrupted a cyberattack where criminals used AI to identify and exploit a zero-day flaw that could bypass two-factor authentication in a widely used system administration tool. The incident is a pretty clear signal that AI-enabled offensive security is moving from theory into actual enterprise risk.

AI gateways manage LLM inference traffic, focusing on cost, routing, and observability. MCP gateways govern agent-to-tool interactions, providing centralized authentication and access control. While both offer infrastructure visibility, neither provides full session-level behavioral context, leaving security teams needing dedicated platforms to detect complex, multi-step agentic attack chains.

Gartner says fully sovereign cloud is only realistic for China or the US, complicating European efforts to reduce dependency on American hyperscalers. The argument matters for IT leaders balancing compliance, national data rules, and practical infrastructure constraints.

Engineering leaders are stuck between "ship AI-generated code faster" and "review every line manually." Snyk's free leader's guide maps a 4-step path to "Secure at Inception" — guardrails embedded in the AI workflows developers already use, so security stops blocking and starts shipping. Get the eBook →

OpenAI agreed to acquire UK-based AI consulting firm Tomoro as it pushes deeper into enterprise deployment and implementation. The acquisition suggests OpenAI wants more control over how companies translate model access into real operational value.

Amazon launched Amazon Supply Chain Services (ASCS) for businesses of all sizes. Companies can now use Amazon's freight, warehousing, fulfillment, and delivery network. Brands like Procter & Gamble, 3M, and Lands' End are already using the service to move and deliver products faster.

Yum Brands is modernizing Taco Bell, KFC, and Pizza Hut around a common data model to support AI and digital operations across 35,000 restaurants. It is a useful example of how “AI transformation” often starts with standardizing old systems and fragmented data.

Enterprise AI agents face tool registry poisoning risks that artifact-based security like SLSA cannot prevent. Behavioral integrity requires a runtime verification proxy to enforce endpoint allowlisting, discovery binding, and output schema validation. Implementing these controls ensures tools act as declared, mitigating risks like prompt injection and behavioral drift.

The European Commission said OpenAI offered to provide access to cybersecurity tools through its EU Cyber Action Plan, while Anthropic has not made a comparable offer yet. The move positions defensive AI access as part of OpenAI's regulatory and public-sector strategy in Europe.

Practical insights from European enterprise leaders about building the conditions where people trust and adopt AI.

The UK's data protection watchdog fined South Staffordshire Water's parent company over security failures tied to a Cl0p ransomware attack.

cURL creator Daniel Stenberg said Anthropic's bug-hunting Mythos model looked more like marketing than a major security breakthrough after it found only one low-severity cURL issue.