Prototype / demo
About this app
Risk Atlas is a public desk for seeing AI safety risk in one place — what is showing up, how bad it might be, what would reduce it, and where the work gets stuck. That matters because the record is scattered across lab PDFs, institute evals, and posts on X, while compute money outruns the people who can test and halt a system. A shared, hostile reading of the residuals is a start. It is not a substitute for a formal safety case.
What the app does
It takes public signals (mostly posts on X), files them on a three-axis taxonomy, and rolls them into overarching risks written in NASA statement form. Each risk has experimental 1–5 scores for consequence, likelihood, and urgency, several failure scenarios, and a residual after mitigations marked in place. Capital and institute budgets are mapped onto those cards as corroboration. Process handoffs name where a control would actually have to sit.
Built independently with Grok by Doug (@risk_atlas). Not an official xAI product, service, or endorsement. 100% vibe coded.
Every page
- Dashboard
One view of the four registers: top residual risks, critical signals, open controls, and where signals cluster on the taxonomy.
- Top 20 risks
NASA-form report cards with consequence, likelihood, urgency, failure scenarios, PRA aspect tags, and a human-review layer. Filter by industry or personal impact for that slice’s own top 20.
- Watch register
Hundreds of NASA-form risks below the global cut, plus a matching 100 for each industry and personal lens.
- Signals
Public posts from X, classified onto public impact, vulnerable systems, and industries. Each card links back to the original post.
- Mitigations
The control book: who should own it, expedited vs normal calendar time, and which risk it is meant to move. Shared across industries.
- Bottlenecks and interfaces
Process seams where residual leaks — eval to gate, lab to institute, product to buyer, incident to disclosure — and the control that would fortify each one.
- Taxonomy
Three axes for filing: how the harm lands, which system fails, and which sector is in the blast radius.
- Sources
OECD AIM, AI Incident Database, OWASP GenAI, and lab RSPs mapped onto the same cards. Originals linked.
- Charts
XY and bar reads of residual seepage, signal clusters, control status, and lab / institute capital.
- Timeline
When signals were injected into the desk — daily rate, severity mix, and source (lab, institute, press).
- Mind map
How the 120 global risks cluster and link.
- Capital
Public-reported VC and grant flows aimed at safety or security, split onto register risks. Frontier compute raises are kept separate.
- Institutes
AISI-network budgets and recent eval disclosures, with an experimental allocation onto the same risks.
- Join
How to read, review, or fund the desk.
How scoring works
Composite = (likelihood × consequence) + urgency. Residual only counts controls marked in place. Aspect tags (capability, domain knowledge, affordance, impact domain) and short pathway notes are a lightweight PRA worksheet — not a causal graph and not a formal PRA report. A reviewer can override a residual with a written rationale; that does not certify the number.
Other public databases
X does not produce two hundred distinct incidents in a day. The extra watch cards are expansions of this sweep’s themes. If you want a second source wired in, pick one:
- AI Incident Database (incidentdatabase.ai)
- OECD AI Incidents Monitor
- AIAAIC (algorithmic controversies)
- MIT AI Risk Repository
- MITRE ATLAS
- OWASP LLM / Agentic Top 10
- NIST / CAISI notes and NVD/CVE
- UK AISI and the International Network of AISIs
- Lab RSPs and system cards
- METR / Epoch evals
Limits
Not certified, audited, or regulatory-grade. Not for operational or compliance decisions. Scores, dollar splits, and sweep notes are experimental estimates and can be wrong or stale. When a signal is from X, we attribute the author and link the original post — this is not a republishing feed.