09/19/2026

Exp 03/19/2027

cypurr@protonmail.com

cypurr.org

Sept 19th

Outline:

Intro
What is AI
AI and Privacy
What to do
Q&A
10 m Break
Turning off AI settings

Ground Rules

We DO NOT tolerate language or behavior meant to demean or harm folks based on their identity

Security is scary, let’s be gentle.

Step up; Step Back

Privacy:

We’re in public! Keep it general.
Stories Stay. Lessons Leave.
Journalists, Researchers, etc make yourselves known!

Intros

Name (Pronouns)

What's your one-sentence review on AI?

What's a question you hope to answer today?

Who are we?

We are Cypurr Collective

Cat enthusiasts hosting beginner-friendly cybersecurity events with a holistic approach. Teaching to build community self-defense

Next open meeting

(info via email list)

10/21 @7pm

Basics of AI -context

1966- ELIZA "chatterbot"

Basics of AI - how it started

Chatbots followed a list of rules written by a human

For a long time, focus was adding all the rules needed for convincing language

Idea of machine teaching itself on pause until 1990s

Basics of AI - How it's goin'

Supervised learning

- "here's the right answer"

Self-supervised Reinforcement learning

- "Try a bunch of things, get a cookie for better attempts"

- Mostly human feedback for popular AI

CyPurr Collective loves____

cats!

beguiled

authority

Basics of AI - LLMs

[Welch Labs - YT]

Most of "AI" these days are Large Language Models (LLMs), predicting one word at a time based on past training

Basics of AI - LLMs

CyPurr

1. Collective

3. is

2. hosts

4. utilizes

...

Autocomplete will rank words as most likely

LLMs are a bit different

Basics of AI - LLMs

CyPurr

Collective

utilizes

hosts

Combines words often seen together

Note: LLMs can't tell the difference between words on each "die"

Basics of AI - LLMs

CyPurr

Collective

loves

will

has

cats

parties

security

nyc

LLMs are good at selecting die based on context

Better LLMs have more "dice" with fewer "sides", and can have more context

Why now?

Basic methods of AI training haven't changed much, but the scale has.

More data -> 🏭 more data processing 🏭-> better results

This is why PCs and game consoles are more expensive

Current state of things:

- Race to own more processing (US companies)

- China & open source: make it require less compute

- Hitting Plateau of capability, using tricks to keep perceived improvements

- Running at a loss to drive adoption (think Uber)

LLMS

AI Agent

Basics of AI - AI today

Company prompt: You are a helpful assistant, make users use more AI

memory:

user owns a cat

user prompt: What is cypurr?

Use xyz tricks to improve user prompts

"reasoning"

Use xyz tricks to improve user prompts

user wants to know something I should check the web

Am I saying

something bad

LLM

web

LLM

answer knowing

websites say X

and....

Basics of AI - LLM chatbots

Results

Practical takeaways

LLMs as "chunky smoothies" of information it's trained on

Rare info likely to be spit out verbatim "chunks"

More unique/new "blended" combinations of common content

"Blackbox", "nondeterministic". Won't know what comes out of the straw.

No sense of reality, will confidently "hallucinate" or pass along misinformation

Privacy impacts

Collection:

LLMs are mostly trained on the "open internet"
- any information you can access without a login
Eyes on novel/private data
- Explicit example: pay-for-data schemes
- sneaky example: your input/context

Processing:

Makes surveillance data easier to (mis)use
Can de-anonymize online speech

Surveillance

Advertising

A.I.

Cops

Privacy impacts

Sharing:

Everything AI sees, the companies see
Generally held for >30 days
Scanned, sometimes shared with law enforcement
often used for training, possibly "baked" into the model and shared with others
AI agents with access to data may accidentally share it

🏭

🖺📷

🏭

AI doom?

LLMs have learned a lot of common hacking methods

They can try many of them very quickly in new contexts

AI companies mess up "sandboxes" in testing

Coordinated fear mongering so people panic so the public...

- put public resources behind the tech ("for safety")

- restrain open models that threaten them

Real AI Doom

Qs?

10 min break

Keeping AI
in Check

Threat model!

Set priorities of what you want to protect and can control

No AI or AI-optional alternatives

Gmail Settings

Can't switch? turn off features

google (iphone):

Settings

Can't switch? turn off features

google (android):

Siri Settings

https://www.eff.org/deeplinks/2026/09/how-limit-what-apples-new-siri-ai-can-access-ios-27

TLDR:

Open Settings > Privacy & Security > Analytics & Improvements
Disable the option for “Improve Siri & Dictation.”

Open Settings > Siri > App Access
Tap an app where you’d like to change Siri’s settings.

Open up Settings > Screen Time > Content & Privacy Restrictions
If you have never done so, enable the toggle for “Content & Privacy Restrictions.”
Tap the Siri option, then “Allowed Siri Version.”
Select “Siri Classic.”