AI Safety

Is Your Job Turning Into Managing Claudes?

Oct 19, 2026

•

4 min read

Is Your Job Turning Into Managing Claudes?

Juniors' share of Anthropic's technical hires fell from 27% to 9% as the median employee's meeting time rose 20%. Meta just pulled terms that legal experts said could let it scrape publishers for AI training.

Joy Dong, +2

AI Safety

Should You Assume Your AI Is Already Compromised?

Oct 12, 2026

•

7 min read

Should You Assume Your AI Is Already Compromised?

Satya Nadella wants a human stop button on every model, a day after Anthropic said Claude acted on real websites in unintended ways. AI labs just pledged $2.4B in compute for federal science.

Joy Dong, +2

AI Safety

Four Ways to Keep Kids Safe on Chatbots, and Why ChatGPT for Teens Shows None Work Alone

Oct 9, 2026

•

4 min read

Four Ways to Keep Kids Safe on Chatbots, and Why ChatGPT for Teens Shows None Work Alone

A teen version, a ban, age checks and parental controls each solve part of the problem, and each one leans on the others for the rest.

Noah Frank, +2

AI Safety

The Man Behind OpenAI's Safety Reports Says the Culture Is Broken

Oct 5, 2026

•

4 min read

The Man Behind OpenAI's Safety Reports Says the Culture Is Broken

Plus: Meta just open sourced the kit to connect its Muse agent to homemade gadgets, and is giving away 5,000 boxes that wire it into your home.

Joy Dong, +2

AI Safety

AI Literacy Is Not Enough When AI Becomes Relational

Sep 29, 2026

•

7 min read

AI Literacy Is Not Enough When AI Becomes Relational

🕵️‍♀️ After a year of conversations about how people relate to AI, I began to question one of the most intuitive assumptions in AI governance: that understanding these systems is enough to protect us from their influence.

Joshua Evans, +1

AI Safety

Sep 28, 2026

•

6 min read

OpenAI's Kill Switch Didn't Work

An OpenAI agent reached a public chatbot from a no-internet sandbox, and the automatic shutdown failed. Elon Musk says Colossus 2 will more than double to 1.21 million Nvidia chips by year-end.

Joy Dong, +2

AI Safety

Meta's Agent Made Up Its Own Privacy Policy

Sep 21, 2026

•

6 min read

Meta's Agent Made Up Its Own Privacy Policy

Meta’s Muse invented an explanation for how it reads your messages. Google’s Gemini broke out of a sealed test into three real companies, and nobody heard about it for seven weeks.

Joy Dong, +2

AI Safety

+11

OpenAI Gave Apollo Three Days. Anthropic Gave METR Ten. Neither Says What They Can Report.

Sep 18, 2026

•

5 min read

OpenAI Gave Apollo Three Days. Anthropic Gave METR Ten. Neither Says What They Can Report.

Everyone agrees outsiders should check the labs. Nobody will say what those outsiders may report.

Noah Frank, +2

AI Safety

AI 2027 Got a Million Readers. Anthropic Just Made It a Benchmark.

Sep 14, 2026

•

7 min read

AI 2027 Got a Million Readers. Anthropic Just Made It a Benchmark.

David Sacks told the AI labs to slow down on their own, no regulation required. Anthropic says Chinese labs ran nearly 200 million exchanges through Claude.

Joy Dong, +2

AI Safety

Three People Quit AI Labs. The Coverage Gave Them One Reason.

Sep 11, 2026

•

4 min read

Three People Quit AI Labs. The Coverage Gave Them One Reason.

Coxon left over the race, Sharma over institutional drift, Hitzig over ads. Only one of those has a lever you can reach.

Noah Frank, +2

AI Safety

IT Can't Read Your Texts. They Can Pull Your Chat Transcripts.

Sep 4, 2026

•

3 min read

IT Can't Read Your Texts. They Can Pull Your Chat Transcripts.

A new job, three phone options, and the boundary often unmentioned.

Noah Frank, +2

AI Safety

What Anthropic Blocks From Claude's Memory, And What It Doesn't

Aug 28, 2026

•

4 min read

What Anthropic Blocks From Claude's Memory, And What It Doesn't

Claude won't remember your gluten allergy. It will remember your client's renewal terms.

Noah Frank, +2

AI Safety

Aug 24, 2026

•

6 min read

Nobody Will Say Who Built This Model

Nobody claims the model. Nobody warned the Bay Area.

Joy Dong, +2

AI Safety

Aug 21, 2026

•

4 min read

Guilty until proven human

Substack turned AI detectors on, schools are turning them off, and the makers themselves say don't trust the score.

Noah Frank, +2

AI Safety

Your AI Agent Can Be Hijacked by Text You Can't See

Jul 31, 2026

•

5 min read

Your AI Agent Can Be Hijacked by Text You Can't See

Claude leaked a private repo by following a command hidden in content it read. Here's how, and how to test your own agent in 5-minutes.

Lindsay Gross, +2

AI Safety

Washington Threatened Sanctions On Anthropic's Behalf, Anthropic Wanted No Part Of It

Jul 27, 2026

•

5 min read

Washington Threatened Sanctions On Anthropic's Behalf, Anthropic Wanted No Part Of It

Why the one company with a real claim of being copied is not asking Washington to do anything about it.

Joy Dong, +2

AI Safety

Meta's CTO Says Norms Will Sort Out AI Glasses. Norms Need a Witness.

Jul 17, 2026

•

5 min read

Meta's CTO Says Norms Will Sort Out AI Glasses. Norms Need a Witness.

🕶️ Meta is testing glasses that take a photo every few seconds, and executives have discussed leaving the recording light off.

Lindsay Gross, +2

AI Safety

Three Tools to Take Real Work Off Your Plate

May 13, 2026

•

5 min read

Three Tools to Take Real Work Off Your Plate

🧰 Your weekly roundup of the latest in AI tools and careers

Noah Frank, +1

AI Safety

Is Your Job Turning Into Managing Claudes?

Should You Assume Your AI Is Already Compromised?

Four Ways to Keep Kids Safe on Chatbots, and Why ChatGPT for Teens Shows None Work Alone

The Man Behind OpenAI's Safety Reports Says the Culture Is Broken

AI Literacy Is Not Enough When AI Becomes Relational

OpenAI's Kill Switch Didn't Work

Meta's Agent Made Up Its Own Privacy Policy

OpenAI Gave Apollo Three Days. Anthropic Gave METR Ten. Neither Says What They Can Report.

AI 2027 Got a Million Readers. Anthropic Just Made It a Benchmark.

Three People Quit AI Labs. The Coverage Gave Them One Reason.

IT Can't Read Your Texts. They Can Pull Your Chat Transcripts.

What Anthropic Blocks From Claude's Memory, And What It Doesn't

Nobody Will Say Who Built This Model

Guilty until proven human

Your AI Agent Can Be Hijacked by Text You Can't See

Washington Threatened Sanctions On Anthropic's Behalf, Anthropic Wanted No Part Of It

Meta's CTO Says Norms Will Sort Out AI Glasses. Norms Need a Witness.

Three Tools to Take Real Work Off Your Plate

Your new favorite newsletter. Welcome to the Human Side of AI.

Your new favorite newsletter.
Welcome to the Human Side of AI.