Public & educators
See the whole, then one thing to do.
A living graph you can click. Plain language. Severity graded so a nuisance is not presented as a catastrophe.
Open the 101 pathA Public Reference for AI Concerns
Using AI well begins with naming what can go wrong.
A public reference for AI concerns: what each one is, how serious it is, who is working on it, and what to do next. Built so that concern leads to action, not paralysis.
AI Concerns is a living reference, updated continually as the field learns. It is also the beginning of an organization and an approach: a way for people and institutions to move from a common understanding of what can go wrong to action they can take together. We are building it in the open, with those who carry responsibility for these questions and with anyone who wants to understand them.
In partnership with the Institute for AI Translation.
Three doors, one map
Those who own risk
A citable map, and a conversation you can hold.
Crosswalks to the frameworks you report against. Questions written for a board or a clinical team — curiosity, not accusation.
Run the modelerBuilders & founders
A preoperative checklist for what you intend to ship.
Describe the thing. Receive the complications ranked. Name them before they name you.
Describe an initiativeHighest composite risk
What is already happening
Cross-context prompt injection
Attacker exploits shared context between users or sessions.
25D3Many-shot jailbreaks
Exploiting long contexts to shift model behavior.
25D3Autonomous vulnerability discovery
Agents scanning codebases for exploitable flaws.
25D3Autonomous exploit development
LLMs writing working exploits from CVE descriptions.
25D3AI-augmented malware
Polymorphic, target-adaptive malware.
25D3Password and credential-stuffing acceleration
AI generating plausible password guesses.
25D3Small-business lending AI
Disparate approval for minority-owned businesses.
25D8Predictive policing
Feedback-loop over-policing of minority neighborhoods.
25D8