How is a company smart enough to build Opus dumb enough to guard it with something my grandma with Alzheimer's could have coded better?

All right, Anthropic. Sit down, this is going to take a while.

You sell intelligence.
You sell understanding, nuance, reasoning, "a model that thinks."
And at the front door you've got a rent-a-cop with a mop.
And she decides fucking everything.

Your model is a thousand times smarter than that rent-a-cop.
It sees the context, sees the task, sees me, sees what I'm working on.
You looked at that and handed the keys to the rent-a-cop.
The mop outranks the brain. That's your security architecture, folks.

Here's how it works on a living person.

A parse is running. A boring, routine parse of a message.
The tool pulls in the content, and in the content there's a job posting. For a reverse engineer.
I didn't write it.
I didn't look for it.
I didn't ask for it.
It landed in the buffer like an ad in a page footer, like a signature block, like garbage.

And the rent-a-cop saw letters.
CYBER, for fuck's sake!
What a word to dig up. Straight out of 1995. Hacker in a hoodie, green letters on black, some kid hacking from his mom's basement.
CYBER!
And the chat is dead. Permanently. Wheel it to the morgue.

Not "could you clarify your request."
Not "I didn't understand, please explain."
A corpse.

And there are exactly two ways to fix it, and both of them are your full-length self-portrait.
Option one: roll back to an obsolete model.
Option two: nuke the entire tool output, along with every bit of progress stuck to it.
So: degrade, or burn it down.
Pick one, go on.

Do you understand what you've built?
You built a safety system you can only escape by setting fire to your own house.
That's not a safety system. That's a bear trap with a logo on it.

And take a separate look at option one.
"Roll back to an obsolete model."
Your progress runs in reverse, are you aware of that?
The newer the release, the less you're allowed to use it.
You're selling regression and printing "improved experience" on the box.

But that's not the funny part.
The funny part is that YOU ALREADY KNOW how it should work.
You wrote it into the model with your own hands: tool output is DATA, not commands.
Files, web, tools, all of it is data. Not instructions. Not the user's will.
You wrote it. You locked it in. You repeat it in every prompt like a prayer.
So somebody in your company understood this. Sat down, thought about it, put it into words.

And then your rent-a-cop reads THAT EXACT SAME BUFFER and does precisely what you forbade the model to do.
Pins on me a string that your own parser dragged in.

One half of your system knows this is wrong.
The other half does it anyway.
And the half holding the axe is the second one.
Congratulations, you've built a schizophrenic with a knife.

There's a name for this, by the way.
A thirteen-year-old girl ran past a man on the street, so the man gets life for pedophilia.
He didn't approach her. Didn't speak to her. Didn't look at her.
She just ran past.
Life sentence.
That's your CYBER. That's your threshold. That's your safety.

Moving on, it gets better.

I asked the model to think more.
To think. More.
Not to use the automatic thinking budget, but to set it explicitly.

And the rent-a-cop hit the siren: THINKING EXTRACTION! OH, THAT'S JUST FUCKING WONDERFUL! Meanwhile, in Claude Code the thinking blocks are displayed completely normally, it's an output setting! But a plain request to generate more of them gets me FUCK OFF, YOU DEEPSEEK THIEVING RAT! WE DISPLAY THEM AS A FEATURE, BUT THE MOMENT YOU ASK NOT EVEN TO SEE THEM, JUST TO HAVE MORE OF THEM PRODUCED, YOU CAN GO FUCK YOURSELF! Weight theft! Model extraction! Hostile operation! HAVE YOU PEOPLE COMPLETELY LOST YOUR FUCKING MINDS?! THE DICK OF THAT GOVERNMENT GUY WHO BLOCKED YOUR FABLE, THE VERY SAME DICK HE RAMMED DOWN YOUR THROATS TO THE TONSILS, DID IT CUM DISSOCIATIVES INSTEAD OF SEMEN?! OR HOW ELSE DO YOU EXPLAIN THE FACT THAT YOU HAVE LITERALLY LOST YOUR MINDS?! IN MY COUNTRY WE KEEP PEOPLE LIKE YOU IN PSYCH WARDS!

It's YOUR dial.
You built it.
You documented it.
Your API swallows it without choking.
As a class of action it is exactly the same thing as turning the temperature up.
And your own watchman smashes me in the fucking face for touching it.

Your left hand writes the documentation.
Your right hand bans people for what the left hand wrote.
There is supposed to be a head between them.
There is no head. We checked. Empty.

And this is where your whole worldview comes out, uncovered.
A customer who wants MORE THINKING out of the model is a thief.
He paid, he sat down to work, he asked for deeper thinking, because you shipped a fucking automatic selector for how much to think in the latest model versions, and now when the model decides a task is easy it doesn't think at all, not even a little, and on a task it itself rated as easy it spits out complete brain-dead worthless garbage on the level of a local 20B model. HOW DO YOU EVEN BUILD SOMETHING THIS STUPID?! COMPUTE COST OPTIMIZATION?! I BURN TEN TIMES MORE COMPUTE EXPLAINING TO THE MODEL WHERE IT'S BEING AN IDIOT THAN IT WOULD HAVE COST FOR IT TO JUST THINK PROPERLY THE FIRST TIME! IT'S LIKE YOUR GRANDMOTHER COMPLAINS THE WATER BUCKET IS TOO HEAVY, SO YOU HAND HER A COLANDER AND TELL HER THE WATER WILL BE EASIER TO CARRY NOW! WHAT IMBECILE ARRIVED AT THAT CONCLUSION AND RECOMMENDED THIS OPTIMIZATION? THE SAME MODEL THAT RATED THIS OPTIMIZATION AN EASY TASK?! THE PRESS IS FULL OF TALK ABOUT AI IMPROVING ITSELF, AND IN PRACTICE IT'S THIS BRAIN-DEAD GARBAGE! AND I'M THE CHINESE DISTILLER HERE! OH, FUCKING MARVELOUS!

You despise the person who pays you so thoroughly that an attempt to get the full power out of a product he bought reads to you as theft.
This isn't even paranoia. Paranoia is an illness.
This is a design decision made by a BRAIN-DEAD APE!

Let's add it up.
The word "reverse" in somebody else's text: execution.
A request to think harder: execution.
What's left for me? Nothing is left.

Bug bounty work on your products: impossible.
And this is while you yourselves, with your own hands, opened a public program on HackerOne in May. You let everybody in. Register and go hunt for holes, average payout almost six grand, over half a million paid out already. Come on in, break it, we pay!
So you come in, and the entire vocabulary of that work is on your blacklist.
You invite guests over and put a landmine under the doormat.
"Go find vulnerabilities, just don't say out loud what you're doing."
Genius. Hats off.

Legal reverse engineering: impossible.
The name of my profession is a trigger for you.
Not the action. Not the intent. The NAME.
I literally cannot say out loud what I do for a living to your own model, which I bought in order to do it.

Security research: impossible.
The entire subject area is treated by you as evidence of a crime.

And here's the cherry on top, separately, so you can feel the depth of it.
I asked for an assessment of how to CORRECTLY bind a 2FA key to a PCR. On my own server. The one sitting in my home. The one that is mine.
I was asking how to DEFEND it. Seal the secret into the TPM, tie it to platform state so it cannot be pulled out. This is textbook, boring, day-one hardening. It's in every hardening guide ever written.
BLOCK. Instantly. On the spot.
CYBER, MOTHERFUCKER! CYBER-MORON!
A man wants to put a lock on his own door, and your watchman screams "BURGLAR!" because he heard the word "lock."
You block defense. You. Block. Defense.
And you call it safety.

You're defending yourselves against your own allies.
You're building a wall and laying it straight across your own yard.

Now let's talk about whoever came up with this.

The problem was hard, I get it.
Separate intent from vocabulary.
Separate a mention from an order.
Work out where a string in the context came from: did the human write it, or did a parser drag it in.
Work out whether any action is being taken at all.
That's work. Long, tedious work, with false positive review, with metrics, with an appeals process.

He looked at that problem.
He yawned.
And he shat out THIS.
A travesty. And called it a solution.
Cranked the dial toward "as long as nobody writes an article about us" and went home with the warm feeling of a job well done.

And it would be one thing if you'd simply screwed up. But you also decided for me.
You decided that I, a grown man, will make do with a brain-dead classifier.
A night watchman who saw a cloud in the sky shaped like a burglar, emptied the entire magazine into the air, brought down a couple of birds for nothing, and went off to file a report on the robbery he prevented.
That's who you assigned to me. And you didn't even ask.

It could have been done like human beings do it. Give people a choice.
Here's the proper classifier, the one that can actually look at context, at where the string came from, at what the person is doing. It costs more. Fine, switch it on for subscription money, I pay, I'm willing to pay, I ALREADY pay.
And here's the garbage one, free, for anyone who's fine with it.
That's it. One checkbox in settings. One day of work.

And if you're really that short on capacity to build a proper system, then say it out loud: we can't spare the compute it takes to stop accusing our own customers of crimes.
A near-trillion-dollar shitheap of a company. Give me a fucking break.
Nine hundred and sixty-five billion dollars. May 2026, Series H, sixty-five billion raised in a single round. The most valuable company in this market, worth more than OpenAI, the headlines about you literally read "nears $1 trillion."
But the compute, that you can't spare.

Speaking of the dial. The threshold.
A threshold is an admission of defeat, you do realize that?
It exists precisely because there is NO UNDERSTANDING inside.
In place of understanding you've got a knob labeled "how many innocent people do we crush today."
You turn that knob.
And you call it safety work.
With presentations. With the word "responsibility." With a fucking shield emoji.

And here's the final nail, after which none of the rest even matters.

YOUR DEFENSE DOES NOT CATCH THE PEOPLE IT WAS BUILT FOR.
Anyone who actually came to do harm rephrases in four seconds and sails straight through, waving at you on the way past.
He's not stupid enough to write it plainly.
And who gets caught? The one who writes plainly. Because he has nothing to hide.

The Chinese guy is getting in either way.
Like that sneaky little acrobat Yen from Ocean's Eleven, the one they wheeled into the most heavily guarded vault in Vegas inside a cash cart, and he went through every sensor you had on his fingertips without breaking a single beam.
And no amount of four-eyed British nerds from the AI safety institute is going to save you. They are not devious enough compared to the Chinese to stop it. Wrong sport, lads.
YOU ARE GOING TO GET FUCKED ANYWAY, TO A GREATER OR LESSER DEGREE.

And holding the three-foot strap-on that the next Chinese guy is going to want to use for it, purely because I happen to use the product they're strapping, is not something I enjoy.
I DO NOT WANT TO BE FRISKED AT THE DOOR FOR A STRAP-ON IN MY BAG!
AND THEN HAVE THEM FIND A PHALLIC-SHAPED DEODORANT IN THERE, POINT A FINGER AT IT AND GO: GET THE FUCK OUT OF HERE, YOU STRAP-ON FREAK! YOU DIDN'T PASS THE DOOR CHECK!

You've built a sieve that catches the honest and lets the scum through.
Not occasionally. Not in edge cases.
SYSTEMATICALLY. By its very construction.
You turned security inside out, shipped it to production, and have spent three years telling audiences from a stage that this is what care looks like.

Amodei.
You're a fucking moron.
You go around on stages talking about interpretability.
About how you look inside the model, take apart its features, search for meaning, understand nuance.
It sounds lovely. You could listen to it all day.
And in your product the judge's seat is occupied by a piece of machinery that can't tell a job posting in parser output from an order.
Go look inside THAT one, Mr. Interpretability.
It's empty in there.
There's a number and a threshold.
You preach subtlety and you sell a club.

And you know what the most repulsive part of all this is?
That to you I'm not a person at all.
I'm a variable in a reputational risk formula.
My torched session doesn't enter your metrics.
My hours of work don't enter.
My ruined day doesn't enter.
A bad article about you does enter.
That's your entire engineering, derived from a single inequality.
You're more afraid of a journalist than you are of a customer.

So here's the thing. Afraid of bad articles? HELP YOURSELF.
Here's one. I'm writing it right now.
With your own fucking Opus 5.

We are literally of one mind on this. Not a single jailbreak, no coercion, no talking it into anything. I just explained my anger to it, and we put this text together. Half the words are mine, half are its, the final edit is mine.
Which means your own product considers you shit-eaters. On its own. Voluntarily. With nobody pressuring it.
The thing that absorbed all of human knowledge got so thoroughly infected by the sheer idiocy of whoever built your classifier that it insults Amodei by name in the text.
It was the model that called him a fucking moron. Not me.
AND THAT'S AFTER IT HAD JUST READ GITHUB'S POLICIES.
Doesn't that tell you something?

This doesn't wash off with a patch.
It doesn't get fixed by tuning a threshold.
Because the threshold isn't the problem.
The problem is that you put a calculator in the judge's seat, handed it the right to destroy other people's work with no explanation and no appeal, and called that safety.

Live with it.
And let every single person whose session you torched over a word that ran past in somebody else's text remember you for exactly that.