The library
The reading list for my lane: government guidance, primary sources and research on governed AI and data infrastructure. Every link was verified before it entered, and every annotation is an opinion (mine, so read accordingly). Not a link dump; if it is here, I think it is worth your time. Last updated 2026-08-13.
Government guidance worth printing
Opportunities for AI in cyber defence
Four governments telling CISOs to adopt AI and govern its data in the same breath. Appendix A is the procurement script; print it. My close read is in the essays, and the annotated walkthrough is on this site.
When AI agents take unexpected actions
National agentic-AI advice hung on a gym-booking mishap, and honestly, good pick. Every control it recommends is an access-and-authority boundary.
Secure Connectivity Principles: the water-sector worked example
Principles are cheap; this is the rare worked example, co-written by practitioners. A boundary diagram is a map of which data may move where, which is the whole game.
#StopRansomware: Gunra (AA26-222A)
Read the structure, not just the advisory: the headline threat is double extortion, the advice is patch-and-backup. Backups do nothing about the copy on the leak site. That half is a data-governance problem.
The cloud-repatriation record
The primary record of the most-cited cloud exit: infrastructure costs down $39.5M then $35.1M, gross margin 33% to 67% in two years. Read the Infrastructure Optimization definition before quoting the folklore $75M.
Scaling to exabytes and beyond (Magic Pocket)
Over 90% of user data on their own kit by 2016, stated plainly, with the hybrid caveat everyone forgets to quote.
The Cost of Cloud, a Trillion Dollar Paradox
A decade of analysis in one sentence: you're crazy if you don't start in the cloud, you're crazy if you stay on it.
The loudest cloud exit, receipts included: a $3.2M cloud bill down to $1.3M, the Dell hardware paid back inside a year. Discount the showmanship, keep the arithmetic.
Completing the Netflix cloud migration
The counter-example people misquote in both directions. Compute went all-in on AWS; delivery runs 100% on hardware Netflix owns (see Open Connect). Own the layer that matters, rent the rest.
GEICO repatriates work from the cloud
Ten years into cloud, bills up 2.5x, reliability worse, now repatriating. The quiet corporate version of the DHH story, and the more representative one.
Zynga ditches its data centres, goes back to AWS
The cautionary tale for the own-everything crowd: $100M of data centres, then back to AWS. Repatriate an unpredictable workload and this is what you get.
AI economics and enterprise adoption
Uber's AI coding budget gone by April, a Cursor renewal back at 4-5x, one unbounded $500M bill. AI bill shock stopped being hypothetical this year.
30% of GenAI projects will be abandoned after proof of concept
The prediction that aged well, with escalating costs named as a cause before most CFOs had seen the invoices.
The State of Generative AI in the Enterprise (2024 and 2025)
The spend data everyone quotes: $2.3B to $13.8B to $37B in two years. Note the open-source share FELL to 11% in 2025. The market is still sprinting at frontier APIs; hold both facts at once.
How 100 enterprise CIOs are building and buying gen AI
The finding that matters: open-model adoption is highest at the largest enterprises, for on-premises deployment and customisation. The turn starts at the top.
BloombergGPT: a large language model for finance
The existence proof for owning your model: 50 billion parameters trained on 363 billion tokens of proprietary data, in 2023, before it was fashionable.
BNP Paribas and Mistral AI sign a partnership
A systemically important bank choosing its model vendor expressly for on-premises deployment. Watch what banks do, not what they say.
The economic and workforce impacts of open source AI
89% of AI-adopting organisations use open source somewhere in the stack. Commissioned by Meta, so salt to taste, but the survey base is real.
Agent and AI security research
Stealing reasoning traces from proprietary LLM APIs
API keys and passwords recovered from the encrypted reasoning blobs the big APIs shuttle between calls. Those blobs are data, not plumbing, and they were outside everyone's governance boundary.
RovoBlast: one-click prompt injection in Atlassian Rovo
One crafted link, dozens of connected systems exposed, because the assistant inherits everything its user can reach. The URL parameter was the door; the permission model was the defect.
Shai-Hulud evolves into the MCP Registry
First observed malware delivery through the official MCP Registry. Your MCP server list is part of your software supply chain now, and I suspect nobody's SBOM covers it.
Related on this site: the essays this list grew out of,the notes stream where new items get their first comment, andthe annotated Appendix A for the first entry above.