The AI Doomers over at the "AI in Context"'s

YouTube channel just released yet another video

to try to scare you about SuperAI. Most of that

video is reading a fictional story out of a

book

that I've talked about several times - I'll link

those videos - or repeating the opinions of

people

in the industry who have a financial interest

in AI being considered world-changing. But

there

are a handful of verifiable claims scattered

here and there, and even with only a tiny,

tiny part of the video involving the real world

at all, they still couldn't help lying to you

to

try to scare you. And there's very little I

hate more than rich and powerful people lying

to the

public - or paying someone to lie to the public

on their behalf - in order to get richer and more

powerful. What a bunch of As-----ls This is the

Internet of Bugs. My name is Carl. I've been a

software

professional since the late 1980s. I'm not

being paid to make this video and I have no

financial

incentive to tell you what I'm about to say. I'm

just trying to do my part to make the Internet

a

safer and less buggy place, because I think

those of us that are able, should be responsible

to try

to make the world better for everyone else. So

here's the claim from the video: "In February

2026, Anthropic asked their most advanced AI

model to autonomously find zero-days. These are

previously unknown software vulnerabilities.

Four zero days were enough to take down Iran's

nuclear program. Their AI, Claude Opus 4.6,

found 500 of them. In a post about this,

Anthropic wrote that, quote, 'the same

capabilities that help defenders find and fix

vulnerabilities

could help attackers exploit them.' Unquote.

Less than a week after that post, a single

hacker

used that same AI model to hack the Mexican

government and steal 195 million taxpayer and

voter records." So the model in video is talking

about is Claude Version 4.6, which was released

in February of 2026. That's the version that

Anthropic claims found 500 zero days, although

we can't verify that claim. And AI's are

notorious for finding security bugs that don't

actually

exist. So here's the paper on Claude helping

that attacker attack the Mexican government.

This paper was published on February 26, 2026,

not long after Claude's 4.6's release.

But here you can see that the hack actually

took place between December 2025 and early

January of

2026, happening months before the "AI in context"

video claimed it had. And the hack didn't

involve

any zero days. The bugs exploited the attack

were: "weak or default authentication

configuration,"

"exposed admin channels", and "web applications

that still had unpatched vulnerabilities from

2023", which means that the people responsible

for securing those servers never bothered to

install fixes that had been available for more

than a year at that point. What happened here

was

that Anthropic trained their model Claude to

know about exploitable bugs that were found in

2023.

And that model used that training that it got

on those bugs to write scripts for a bad guy so

that he could hack neglected and out of

date servers. And now this YouTube video is

trying to

convince you that means that a future super AI

will be able to use super zero day hacking

skills

to pursue goals of its own that are harmful to

humanity. This is typical Doomer propaganda.

The reason they want to scare you about AI and

are willing to lie to do it is because as long

as

the public can be convinced that super AI will

kill everyone, and as long as the public can be

convinced that the Chinese can't be stopped

from making super AI, then Silicon Valley could

do whatever

it wants in the service of getting to super

killer AI before the Chinese do. I've got a

more detailed

write up on this on my SubStack over here. Now,

why would 80,000 hours - the group who produces

the videos for the "AI in context channel" - want

Silicon Valley to be able to do whatever it

wants?

Well, I don't know for sure, but I have a guess.

So 80,000 hours' biggest funding source is a

group

called "Coefficient Giving", which was started by

Dustin Moskovich, who was one of the lead

investors

in Anthropic and an early backer of OpenAI. So

the fewer liabilities that Anthropic and Open

AI

have, the more they will be worth, and the more

that they are worth, the more the investments

of

the biggest funder of 80,000 hours is worth.

This is a scam to try to keep the public from

worrying

about: "AIs that encourage troubled teens to

commit permanent self harm", "send the wrong

people to jail

because of bad facial recognition", "modify

pictures of underage girls to make them appear

to be nude",

and were potentially involved in the selection

of an Iranian elementary school to be targeted

by two US Tomahawk missiles. Don't let them

distract you. And if you're going to be scared

of AI,

be scared for the people it's harming right now,

and not for some hypothetical super AI that

will

probably never happen. Thanks for watching, and

let's be careful up there.