101 · Where to start

From anxiety to a question you can actually ask.

You do not need to master 1,700 risks. You need a way in that matches who you are, and a sentence you can take into a room.

• Just arriving

  I want to understand this

• Household

  I am a parent or caregiver

• Health ecosystem

  I am a doctor, nurse, or clinician

• Building in health

  I am building or bringing a tool into care

• Community

  I want to start a conversation

• In the room

  I am a patient, or I care for someone who is

1. 01

   Name the problem.

   The world will not lack AI. It will lack people who can use it well in a clinic, a home, a school. Naming what can go wrong is how you use it well.

2. 02

   Not every problem is the same size.

   Each concern has a score from 1 to 25. Severity is how bad the harm is. Likelihood is how often it already happens. A nuisance is not presented like an irreversible harm.

3. 03

   Talk before you decide.

   Start with a conversation. Seek understanding. Listen until you could restate. Leave with a question. Then pick a path that matches who is in the room.

4. 04

   Questions before verdicts.

   Ask is written for a patient, a clinician, a nurse, a parent, a person building in care. Each path ends in questions you can take into a room.

5. 05

   One thing this week.

   Pick a door. Hold one question with someone else. If you work in care, open Health. Come back when the news changes.

Six problems people are already living with

• Cross-context prompt injection

  Cross-context prompt injection occurs when instructions from one conversation, user, or shared workspace improperly influence another.

• Many-shot jailbreaks

  Many-shot jailbreaking tries to shift a model's behavior by filling its context with many examples of the behavior the attacker wants.

• Autonomous vulnerability discovery

  Autonomous vulnerability discovery uses an AI-driven system to search for software weaknesses with limited step-by-step human direction.

• Autonomous exploit development

  Autonomous exploit development uses an AI agent to turn information about a software flaw into a way to take advantage of it.

• AI-augmented malware

  AI-augmented malware is harmful software whose creation, modification, or operation is assisted by AI.

• Password and credential-stuffing acceleration

  AI-assisted password attacks use learned patterns to improve guesses, while credential stuffing reuses passwords already stolen from another service.