It started as a small thing: a reader — well, me, reading my own site — noticed nine days in July with headlines but no “From the Trenches” commentary underneath them. Easy enough to backfill by hand. But backfilling isn’t fixing, and I’ve learned the hard way that a bug you patch around instead of root-causing just shows up again wearing a different date. ...
Most people talk to an LLM the way they’d talk to a search engine: type a few words, hope for the best, complain when the answer is generic. That’s not a model problem. It’s a prompting problem. ...
Every token an LLM API processes has a price tag, on the way in and on the way out. That fact turns “write better prompts” from a productivity tip into a line item. Most organizations treat AI spend as an unavoidable cost of doing business; in practice, a meaningful chunk of it is just badly engineered prompts paying for their own inefficiency. ...
The perimeter-based defense model — a centralized, physical firewall guarding a well-defined network edge — is effectively obsolete. As business operations move to hybrid, multi-cloud environments, the “network boundary” now stretches to wherever an authorized user or workload happens to be running. ...
Part 1 traced the firewall from stateless packet filters through NGFWs — four generations, each one built to close the previous generation’s blind spot. This part covers what happened once there was no longer a single perimeter to defend at all. ...
Part 1 covered the firewall’s evolution through NGFWs, and Part 2 covered the shift to SASE and Zero Trust. This part is where AI enters the picture — and it enters from two directions at once. ...
Part 3 covered the AI Security Gateway conceptually — inbound, RAG-context, and outbound protection for GenAI traffic. This part is the actual build: a working deployment using LiteLLM, an Apache 2.0 open-source AI gateway, with Redis for state tracking and rate limiting. ...
This series has traced the firewall from stateless packet filters (Part 1) through SASE and Zero Trust (Part 2), the two-front AI security problem (Part 3), and a working AI Gateway deployment (Part 4). This last part looks at where the architecture goes from here — two directions, moving on very different timelines. ...
Every AI model is, at its core, a direct product of its training data. Long before a system answers its first prompt or makes its first real-world prediction, decisions about data collection, sourcing, and processing have already permanently shaped its behavior. ...
In Part 1 we looked at how little organizations actually know about where their AI training data comes from. This time, we look at what happens once that unverified data gets baked directly into a model. ...