Guides
Every platform in this fleet exists because of a specific way AI products fail in production. These guides are the written version of what building — and operating — those tools taught us: the law as it actually applies, the attacks as they actually arrive, and the working routines we'd hand a team shipping AI today. No summaries of other people's summaries; each guide is written from the operator's side of one problem.
This library opened on 1 November 2026. Every guide is dated; when a guide is revised, its date changes.
The transparency duties that reach ordinary product teams — chatbots, generated media, public-interest text — with three worked examples and the evidence trail that proves compliance when an auditor asks.
Read the guidePublished 1 November 2026
Direct, indirect, tool-call hijack and egress exfiltration — the four attack shapes, one fully worked indirect injection, and the structural defenses (least-privilege tools, egress control, approval gates) that reduce injection to a managed risk.
Read the guidePublished 1 November 2026
Vendors rewrite the contract under your product with a web-page edit. The five clauses that actually matter — training use, retention, subprocessors, hidden model changes, liability — and a monitoring workflow a small team can really run.
Read the guidePublished 1 November 2026
Typosquats, combosquats and homoglyphs feed most brand phishing. How the fakes are generated, how certificate-transparency watching spots them before the campaign, and the evidence pack that makes registrar and host takedowns move in hours.
Read the guidePublished 1 November 2026
What to record for every agent action, how SHA-256 hash chaining makes a log tamper-evident — walked through a three-entry example, edit and all — and how replay and incident review work when the trail can be trusted.
Read the guidePublished 1 November 2026
Announced retirements, alias drift and silent behaviour changes: how models change underneath a product, how golden-set evals catch it before users do, and the pinning/fallback drill that turns a forced migration into routine maintenance.
Read the guidePublished 1 November 2026
Each guide covers the problem one of our platforms handles in production, and links to it where the tool genuinely helps — the guide stands on its own if you never click. The fleet's premise is one tool per failure point; this library is the same premise in prose: one problem per page, covered to the depth we'd want if we were solving it for the first time. Start with the failure that is closest to your launch date.
Questions, corrections, or a problem we should write up next: the contact page reaches the team that wrote these.