Two security stories this week highlight how privacy tools aren’t as private as we think, and how AI access might get harder to come by. Both have real implications for businesses trying to stay secure and competitive.

VPN Exit IPs Become Identity Markers

Mullvad, one of the most privacy-focused VPNs, has a fingerprinting problem. Researchers found that Mullvad’s exit IP addresses are surprisingly identifying — they can be used to track users across sessions and websites.

The issue isn’t Mullvad-specific. It’s about how VPN infrastructure works. Most commercial VPNs rotate through predictable IP ranges. Combine that with timing analysis and connection patterns, and you get a unique fingerprint that follows users around.

This matters for business security. Many companies rely on VPNs for remote work and assume they provide anonymity. They don’t. If you’re using VPNs for sensitive business intelligence, competitor research, or protecting trade secrets, you’re more traceable than you think.

The fix isn’t simple. You need infrastructure that truly randomizes exit points and timing. That usually means building custom solutions or using multiple VPN providers with rotation logic — exactly the kind of infrastructure automation we build for clients who need real operational security.

Economic Barriers May Limit AI Access

A new analysis suggests access to frontier AI models will soon be constrained by economics and security, not just technical capability. The argument: as AI gets more powerful, the costs to train and run these models will create natural access barriers.

The economics are already shifting. Training frontier models now costs hundreds of millions. Running them at scale requires massive infrastructure investments. Only companies with serious capital or cloud budgets will have access to the latest capabilities.

Security constraints add another layer. Governments are already discussing export controls on AI chips and models. If you’re building AI-dependent products, you could face supply chain restrictions that have nothing to do with your technical needs.

What this means practically: businesses need to plan for a world where the best AI isn’t freely available. You’ll need to extract maximum value from smaller, more accessible models. That means better prompt engineering, smarter fine-tuning, and custom agents that do more with less.

This is where specialized AI development becomes crucial. Instead of relying on the biggest models, you build systems that combine multiple smaller models, use retrieval-augmented generation effectively, and create agents that solve specific business problems without needing frontier-level AI.

The Infrastructure Reality

Both stories point to the same reality: the tools everyone assumes work don’t always work as advertised. VPNs don’t guarantee anonymity. Frontier AI won’t stay universally accessible.

Smart businesses prepare for constraints. They build security infrastructure that actually protects sensitive operations. They develop AI systems that work with whatever models are available and affordable.

The companies that thrive will be the ones that build robust, adaptable infrastructure now — before the constraints kick in.