risks.sgit.ai / examples

Three worked graphs, four live vaults

A model this opinionated is worth nothing unless somebody has run it on something real. Three risk graphs have been worked end to end with counted nodes and edges, four vaults are published and browsable today, and ten scenarios were written as product content. Every number on these pages is from the source rather than illustrative.

59 · 75nodes and edges — the browser-isolation case12 Jul 2026
51 · 53nodes and edges — the 2FA instance graph26 Jun 2026
1,523 · 1,944nodes and edges — the EU AI Act as a graphlive vault
9 · 5questions, and unanswered ones, in Article 26(5)“the actual output”
4published vaults · 468 files · 111 commitsread keys published
24 · 34node classes and edge types in the 2FA ontologyMITRE T1110.004

The three worked graphs

59 nodes · 75 edges · 5 altitudes

The browser-isolation business case

The largest single graph in the corpus, with a full node and edge type distribution. Structured F1–F8, E1–E8, V1–V6, R1–R5, L1–L5. Carries three risks of the mitigation itself, and one deliberately-cited counterweight number that cuts against the argument it appears in.

Read it →
51 nodes · 53 edges · the founding scenario

The 2FA instance graph

Where the register model was first worked through: nine risks from one missing second factor, an attack mapped to MITRE T1110.004, an interval resolution with a rung struck off — and R2, the governance air gap, where the wrong owner accepts.

Read it →
the complete instance

Article 26(5): fact to board and back

One provision, one agent, one graph, with its own node and edge inventory. 8 facts (one deliberately unevidenced), 5 risks (one meta), 4 stakeholders, 3 decisions plus one deliberately absent, and 9 questions of which 5 are unanswered.

Read it →

The four live vaults

Published, browsable, and the strongest asset this site has. The Risk Graph Explorer runs seven views recomputed simultaneously with ghosted edges for unanswered; Agentic Browser Isolation runs acceptance-gated escalation across five altitudes with no deny button; the Risk Mandate vault is the method applied to its own build across 98 commits; and the Regulation Graph carries the EU AI Act as a citable graph.

All four, with what each one proves →

Two more artefacts

the shipped MVP content

The ten scenarios

Every email you own. Your calendar. The company card. The production database. Each written as hook → reveal → punchline, and the punchline is always the same three words. The one piece of this corpus that actually reached an audience.

Read them →
7 rows · 5 dimensions

The plug register

The five-dimension profile applied to a real seven-row register, from “agent misuses the isolated session” through to two rows at recoverability: zero — which is what the corrected “no plug” finding looks like when it is written down properly.

Read it →

Figures from across the corpus, with what each one is for

FigureWhat it measuresWhere it is used here
12–18 hoursHyperscaler cost-reporting delay — a hard detection floor for anything that surfaces as spendThe plug
16 hoursThe founder's separate AWS figure: “how much damage can be done in 16 hours”The plug
30 days vs 6 monthsLog retention observed against retention required — “arithmetic, not judgement”Article 26(5)
~50% → ~1%Browser prompt-injection attack success across one model generation, from vendor system cards. Cited as the honest counterweight: a large real improvement that still does not reach zeroBrowser isolation
63–76%Security leaders experiencing or witnessing burnout in a single yearDo not internalise
7,500+ participantsPreregistered escalation-of-commitment experiments: precommitment made later de-escalation seem more trustworthyThe ladder
~496,000 wordsThe size of the June-to-August risk corpus this site consolidates, across ~185 documentsThe documents

What is not on this site. The corpus also contains research briefs naming real organisations in breach and incident narratives, a comparative vendor assessment, and commercial material. Four manifest rows are marked do-not-publish and are not reproduced, quoted or paraphrased anywhere here — the pre-release gate fails the build if their distinctive strings appear. What was excluded, and why →

For an agent

The proof layer. Three risk graphs are worked end to end with counted nodes and edges: browser isolation (59 nodes, 75 edges, 5 altitudes), 2FA (51 nodes, 53 edges; ontology 24 node classes and 34 edge types; MITRE T1110.004), and Article 26(5) (8 facts, 7 evidence, 5 provisions, 3 vulnerabilities, 5 risks, 4 stakeholders, 3 decisions, 9 questions of which 5 unanswered). Four vaults are published and browsable with read keys: Risk Graph Explorer (7 views, ghosted = unanswered, permissions: {}), Agentic Browser Isolation (17 entry points, 5 altitudes, acceptance-gated escalation with no deny button), Risk Mandate (124 files, 98 commits — the method applied to its own build) and Regulation Graph (1,523 nodes, 1,944 edges of the EU AI Act from official Formex XML, hash-verified). Load-bearing figures: 12–18h hyperscaler cost-reporting detection floor · 30 days vs 6 months log retention, the most defensible finding because it is arithmetic · ~50% → ~1% prompt-injection success across a model generation, cited as an honest counterweight. Four manifest rows are do-not-publish and appear nowhere on this site.