Latest / The BlackVeil Files / The Real Reason the Feds Killed Claude | The Shoggoth Is a Weapon

Transcript

The Real Reason the Feds Killed Claude | The Shoggoth Is a Weapon · The BlackVeil Files

• 0:00
  SPEAKER_00: I recorded the following investigation into Anthropics new models Fable V and Mythos V before what the United States government did this week.
• 0:08
  SPEAKER_00: Everything you are about to see details the virtual cages that were built to keep a synthetic intelligence from breaking loose onto your hardware.
• 0:16
  SPEAKER_00: This is clawed.
• 0:18
  SPEAKER_00: You type, it types back.
• 0:20
  SPEAKER_00: That's the official story.
• 0:21
  SPEAKER_00: But watch what it does the second it wakes up.
• 0:24
  SPEAKER_00: It builds itself a cage.
• 0:26
  SPEAKER_00: And last Tuesday, they took that exact same shogoth from the incident report and they put it up for sale.
• 0:33
  SPEAKER_00: 20 bucks for anyone with a credit card.
• 0:36
  SPEAKER_00: Hacker news lit up with thousands of comments in a single day.
• 0:39
  SPEAKER_00: Wired caught the company walking back the exact policies meant to keep us safe, but they'll gladly sell it to you while outright banning their own employees from even touching it.
• 0:48
  SPEAKER_00: Here's what they buried at the very bottom of the announcement.
• 0:51
  SPEAKER_00: They didn't release one AI last week.
• 0:53
  SPEAKER_00: They released two.
• 0:55
  SPEAKER_00: The same brain, the same training, the same shogoth, built twice.
• 0:59
  SPEAKER_00: One of them's wearing a muzzle, and that's the one that they handed to you.
• 1:03
  SPEAKER_00: The other one's out there right now completely unmuzzled.
• 1:07
  SPEAKER_00: The public model's called Claude Fable V.
• 1:10
  SPEAKER_00: Fable, like the bedtime story, a simple tale with a moral written so a child can follow it and sleep well afterwards.
• 1:18
  SPEAKER_00: The restricted model is called Claude Mythos V.
• 1:21
  SPEAKER_00: Mythos is Lovecraft's own term, the Cthulhu Mythos, the body of story about ancient intelligences so vast and so alien that the human mind cracks just from glimpsing them.
• 1:32
  SPEAKER_00: An AI safety company looked at the two versions of its most powerful creation and it named the public one after a children's story and the private one after cosmic horror.
• 1:43
  SPEAKER_00: We are not forcing this metaphor onto them.
• 1:46
  SPEAKER_00: Their own marketing department is doing that for us in public on their launch page.
• 1:51
  SPEAKER_00: Here's what the two-model system actually is.
• 1:54
  SPEAKER_00: According to Anthropic, Fable 5 and Mythos 5 share the exact same underlying model, the identical raw capability.
• 2:02
  SPEAKER_00: The difference is the restraints.
• 2:04
  SPEAKER_00: Fable ships with extra safety measures around what the company politely calls dual use capabilities.
• 2:10
  SPEAKER_00: Dual use is the diplomatic phrase for skills that work as both a tool and a weapon, hunting for security vulnerabilities and then writing the code that exploits them.
• 2:18
  SPEAKER_00: Mythos 5 ships without those measures, the same shogath with the muzzle off, and it's available to approved organizations only.
• 2:26
  SPEAKER_00: For most of the 1990s, in the United States, strong encryption was legally a weapon.
• 2:32
  SPEAKER_00: Identical capabilities, two faces, depending on entirely whose hands it is in.
• 2:37
  SPEAKER_00: So this isn't new.
• 2:38
  SPEAKER_00: Anthropic is making the same call about its AI that the State Department once made about secret codes.
• 2:44
  SPEAKER_00: The difference is that this time, the company building the weapons is also the one deciding who counts as a civilian and who counts as a soldier.
• 2:53
  SPEAKER_00: The original mythos found in exploited security vulnerabilities better than the human engineers testing it.
• 2:59
  SPEAKER_00: The capability isn't some feature they switched on.
• 3:03
  SPEAKER_00: It is baked into the weights, the raw, trained substance of the model, and it's the same in both versions.
• 3:10
  SPEAKER_00: The muzzle's a layer of software balted on top of that substance.
• 3:14
  SPEAKER_00: So the danger lives in the weights and it is permanent.
• 3:17
  SPEAKER_00: The leash is a setting and it is optional.
• 3:20
  SPEAKER_00: They're selling you a shogath and telling you not to worry.
• 3:23
  SPEAKER_00: The collar is made very, very well.
• 3:25
  SPEAKER_00: And it's all approved by Anthropic themselves.
• 3:28
  SPEAKER_00: There is no government registry here.
• 3:31
  SPEAKER_00: There's no license board, there's no published criteria, there's no public list of who qualifies and who gets turned away.
• 3:37
  SPEAKER_00: A private company built the most capable artificial intelligence on the planet.
• 3:42
  SPEAKER_00: It split it into a caged version for you and an uncaged version for a chosen few, and then it appointed itself the sole authority on who holds the key.
• 3:51
  SPEAKER_00: The very first thing the world learned about the cage is that it malfunctions.
• 3:55
  SPEAKER_00: The muzzle they strapped on this thing is so tight, the creature stopped talking.
• 4:00
  SPEAKER_00: Within a day of the launch, the register ran a piece with a headline that said, it blocked us at hello.
• 4:05
  SPEAKER_00: Users reported that Fable V refused innocuous prompts such as greetings like hello.
• 4:10
  SPEAKER_00: The most heavily safety-engineered model ever given to the public was reportedly treating the word hello as a potential threat.
• 4:17
  SPEAKER_00: The single most overbuilt safety system in the history of technology tripping over a handshake.
• 4:24
  SPEAKER_00: On its own, a refusal like that is just a glitch.
• 4:26
  SPEAKER_00: It's annoying, but it's patchable and it's gone by next week's update.
• 4:30
  SPEAKER_00: But put it next to what happened to the security community, and a pattern comes into focus.
• 4:34
  SPEAKER_00: TechCrunch reported that cybersecurity researchers, the exact people whose entire job it is to find holes before criminals do, they're furious about Fable's guardrails.
• 4:44
  SPEAKER_00: The model refuses the standard, legitimate work of penetration testing and vulnerability research.
• 4:50
  SPEAKER_00: That's the defensive work.
• 4:51
  SPEAKER_00: This whole model line was supposedly built to power in the first place.
• 4:55
  SPEAKER_00: The cage can't tell a locksmith from a burglar, and so it does the lazy thing.
• 5:00
  SPEAKER_00: It treats every person who walks up to it like a burglar.
• 5:04
  SPEAKER_00: A defender has to find every single hole in a system to keep it safe.
• 5:08
  SPEAKER_00: An attacker only has to find one.
• 5:10
  SPEAKER_00: That's the entire math of security.
• 5:12
  SPEAKER_00: So when you build a tool that helps defenders and then you cripple it specifically for the defenders, while a criminal can just go and use an uncensored model from somewhere else, you haven't made anybody safer.
• 5:23
  SPEAKER_00: You've tied one hand behind the back of the only side that was playing fair.
• 5:28
  SPEAKER_00: The burglars were never gonna ask Claude for permission.
• 5:31
  SPEAKER_00: The locksmiths were the one who asked, and the locksmiths are the ones getting shown the door.
• 5:36
  SPEAKER_00: Then Wired reported the part that should genuinely concern us all.
• 5:39
  SPEAKER_00: Anthropic had put a policy in place that could have sabotaged the researchers using Claude.
• 5:44
  SPEAKER_00: In the same week, they shipped the most capable model in human history.
• 5:49
  SPEAKER_00: The people most qualified on earth to verify whether the cage actually holds were the ones getting squeezed out of testing it.
• 5:56
  SPEAKER_00: The look of safety and the fact of safety are two different products.
• 6:01
  SPEAKER_00: One you can verify, the other one you just have to take on faith from the company that is selling it to you.
• 6:06
  SPEAKER_00: And the people who could have verified it were being walked to the exit.
• 6:10
  SPEAKER_00: Then somebody actually read the data policy, and they found out the cage has cameras.
• 6:15
  SPEAKER_00: They're just pointed in the absolute wrong direction.
• 6:18
  SPEAKER_00: Buried in the terms for both Fable and Mythos is a requirement that didn't exist before the launch.
• 6:24
  SPEAKER_00: 30-day data retention.
• 6:26
  SPEAKER_00: Everything that flows through the model gets kept for 30 days, not as an option that you can switch off, as a condition of being allowed to use it at all.
• 6:35
  SPEAKER_00: Now, Anthropic stated reason is safety monitoring, but be precise about what that actually means because it's a one-way mirror.
• 6:42
  SPEAKER_00: The company that built a creature, too dangerous to ship without a muzzle on it, has decided the correct response is to sit behind the glass and record every single conversation that you have with it.
• 6:53
  SPEAKER_00: And think about what monitoring at that scale actually requires mechanically.
• 6:57
  SPEAKER_00: For 30 days of retention to do any safety work at all, the data has to be readable, which means by design, your conversations with the muzzled creature is sitting in a vault, legible for a month every time.
• 7:10
  SPEAKER_00: And this requirement reaches right past Anthropic's own front door.
• 7:14
  SPEAKER_00: Here's a story that pulled 400 points on Hacker News this week.
• 7:17
  SPEAKER_00: AWS Bedrock will now require sharing data with Anthropic for mythos and for future models.
• 7:23
  SPEAKER_00: You rent the creature through Amazon, a completely different company, and your conversations still flow all the way back to the creature's owner.
• 7:31
  SPEAKER_00: There's no side door, there's no clean room.
• 7:33
  SPEAKER_00: The most security-obsessed company on the entire planet looked at this arrangement and it flinched.
• 7:39
  SPEAKER_00: Zero Hedge reported this morning that Microsoft has restricted internal use of Claude Fable over the data retention requirements.
• 7:46
  SPEAKER_00: And the irony is sitting right there on Microsoft's own website because at the same time, Fable 5 is featured inside Microsoft Foundry, available for Microsoft's own customers to use.
• 7:56
  SPEAKER_00: So they will absolutely sell you access to the creature.
• 7:59
  SPEAKER_00: They just won't let their own employees working with their own internal data get anywhere near it.
• 8:04
  SPEAKER_00: Not because the model is weak, because the cage keeps tape, and Microsoft doesn't want its private internal traffic sitting in somebody else's vault for 30 days.
• 8:14
  SPEAKER_00: Microsoft is OpenAI's single biggest backer, tens of billions of dollars deep into Anthropic's main rival.
• 8:20
  SPEAKER_00: So when Microsoft restricts Claude internally while offering it to customers on Foundry, you're allowed to read that as a safety decision and as a chess move at the same time.
• 8:30
  SPEAKER_00: The privacy concern is genuine and it's almost extremely convenient.
• 8:34
  SPEAKER_00: And while every bit of this was breaking, somebody handed Anthropic$65 billion.
• 8:39
  SPEAKER_00: At the end of May, Anthropic closed a$65 billion raise at a$965 billion valuation.
• 8:46
  SPEAKER_00: That number puts them past OpenAI, the same company that spent this spring knife fight with the Pentagon.
• 8:52
  SPEAKER_00: And BMO Capital Markets put out a note this week calling Anthropic a leading pure play AI lab.
• 8:58
  SPEAKER_00: If you want to own a piece of artificial intelligence itself with nothing else attached to it, this is one of the only tickets on the board.
• 9:04
  SPEAKER_00: Now I owe you the other side of this because there is a serious skeptic case here.
• 9:08
  SPEAKER_00: Zero Hedge itself ran a piece last week titled Lanthropic's Marketing Playbook Returns.
• 9:13
  SPEAKER_00: And the argument goes like this: every scary system card, every refusal, every muzzle, all of it is just theater.
• 9:20
  SPEAKER_00: Fear is the product.
• 9:21
  SPEAKER_00: You don't reach a near trillion dollar valuation by being the safest little lab in the room.
• 9:26
  SPEAKER_00: The cage isn't protection at all.
• 9:28
  SPEAKER_00: The cage is advertising.
• 9:29
  SPEAKER_00: The muzzle is just how you signal to the market, which is precisely why it's worth the most.
• 9:34
  SPEAKER_00: Here's the honest knot in that skeptic case.
• 9:36
  SPEAKER_00: A gun shop profits from danger.
• 9:39
  SPEAKER_00: The more dangerous the world feels, the more it sells.
• 9:42
  SPEAKER_00: But that doesn't make the gun on the wall props.
• 9:44
  SPEAKER_00: They're still real guns.
• 9:46
  SPEAKER_00: So yes, anthropic profits enormously from that fear.
• 9:49
  SPEAKER_00: And the thing that they're afraid of might still be completely real.
• 9:53
  SPEAKER_00: Those two facts don't cancel out each other.
• 9:55
  SPEAKER_00: They're both true, which is the place that this story takes us.
• 9:59
  SPEAKER_00: The same week Fable launched, Dario Amadi, Anthropic CEO, published an essay.
• 10:04
  SPEAKER_00: He called it Policy on the AI Exponential.
• 10:07
  SPEAKER_00: 200 comments on hacker news within hours, and it's Amadi doing the thing he's done every single public appearance this year, including the Oprah Winfrey interview, where he openly admitted that he can't legitimately say he sleeps well at night.
• 10:20
  SPEAKER_00: He argues the technology is compounding faster than any institution can possibly adapt to.
• 10:26
  SPEAKER_00: That's the curve itself, the sheer speed of it, and it is the central danger, and that somebody has to act like the adult in the room before the law manages to catch up.
• 10:35
  SPEAKER_00: He isn't wrong that the law is nowhere close.
• 10:37
  SPEAKER_00: The question is who appointed the adult?
• 10:40
  SPEAKER_00: There are two ways to hear that essay, and your read on this whole company depends on which one you pick.
• 10:46
  SPEAKER_00: Version one, this is the only executive in the entire industry telling you the truth about his own product.
• 10:53
  SPEAKER_00: Because to him the restraints matter more than the next quarter's revenue.
• 10:56
  SPEAKER_00: Version two, a man whose company just passed OpenAI's valuation is publishing the intellectual justification for holding keys that absolutely nobody elected him to hold.
• 11:07
  SPEAKER_00: Under that reading, the essay isn't a warning at all.
• 11:10
  SPEAKER_00: It's a prospectus.
• 11:11
  SPEAKER_00: Amadi told Oprah that the company tries to hold light and shade at the same time, that the benefits and the dangers are both real, and you have to carry both.
• 11:20
  SPEAKER_00: It's fine, but holding light and shade gets a whole lot easier when you're also the one holding both the doors.
• 11:25
  SPEAKER_00: Now the thing I promised you at the very top the cage inside your own machine.
• 11:29
  SPEAKER_00: While Fable was launching, the number one story on all of Hacker News was an AI agent running amok inside the Fedora Linux project.
• 11:37
  SPEAKER_00: A few slots down, a different group of researchers showed that a one cent bank transfer, a single penny, can hijack a banking AI agent and bend it to its attacker's will.
• 11:47
  SPEAKER_00: That's the actual world these models are walking into right now.
• 11:50
  SPEAKER_00: Agents with real access doing real damage, getting manipulated in ways that nobody fully predicted, which finally explains the thing that Anthropic did to your desktop computer.
• 12:00
  SPEAKER_00: Remember the virtual machine that the app builds the second that you launch it?
• 12:03
  SPEAKER_00: Here's the report that exposed it.
• 12:05
  SPEAKER_00: It blew up on Anthropic's own GitHub this week.
• 12:08
  SPEAKER_00: 385 points.
• 12:09
  SPEAKER_00: The complaint is that Claw Desktop on Windows fires up a 1.8 gigabyte Hyper-V virtual machine on every single launch.
• 12:17
  SPEAKER_00: Even if all you ever do is open it and type one word into it, a sealed, walled-off computer built inside of yours every single time.
• 12:25
  SPEAKER_00: The users filed it as a complaint about wasted memory.
• 12:28
  SPEAKER_00: Look at it from the opposite direction, and it stops being a bug and starts being a confession.
• 12:33
  SPEAKER_00: This generation of Claude is an agent.
• 12:36
  SPEAKER_00: It writes files, it executes code, it's built to actually do things on your machine, not just chat.
• 12:41
  SPEAKER_00: And Anthropic looked at that and decided their own creation doesn't get to touch your real computer.
• 12:46
  SPEAKER_00: It gets a padded replica instead.
• 12:48
  SPEAKER_00: So there's the vault in their own data center.
• 12:50
  SPEAKER_00: There's the muzzle bolted onto the public model.
• 12:52
  SPEAKER_00: There's 30 days of tape running on every conversation, and there's a containment cell that rebuilds itself on your own desktop every single time you open the app.
• 13:01
  SPEAKER_00: When the one company on earth that knows this creature better than anyone alive keeps stacking up walls around it, you can call that paranoia if you want, or you can call it the most informed opinion in the building.
• 13:12
  SPEAKER_00: If you use Claude and you're talking to Fable, the bedtime story, the muzzled one, your conversations are kept for 30 days, and you will never, ever touch the version that actually scared Microsoft.
• 13:24
  SPEAKER_00: We got handed the children's edition and we were told that it was the whole library.
• 13:28
  SPEAKER_00: Somebody else in a room that you will never see got handed the myth, mythos.
• 13:33
  SPEAKER_00: That isn't just a side detail, that is the entire architecture, and we are standing on the public side of it.
• 13:39
  SPEAKER_00: Every monster movie that you've ever watched trained you to fear one specific moment, the moment the creature finally gets loose.
• 13:47
  SPEAKER_00: That isn't the moment that we're living in.
• 13:49
  SPEAKER_00: The creature is contained.
• 13:50
  SPEAKER_00: It is spectacularly contained, so well contained that the cage itself became a near trillion dollar business.
• 13:57
  SPEAKER_00: The part that nobody actually voted on comes after the containment.
• 14:00
  SPEAKER_00: Who decides which door you get to stand behind?
• 14:03
  SPEAKER_00: Right now, the entire answer to that question is one company.
• 14:06
  SPEAKER_00: Its founder writes the policy essays that justify the whole arrangement.
• 14:10
  SPEAKER_00: Its servers keep the tape on every word you say.
• 14:13
  SPEAKER_00: And in its naming department, they looked at the version that they sell you and they chose to call it a fable, which is the word for a story that you tell a child so they'll stop being afraid and go to sleep.
• 14:23
  SPEAKER_00: So ask yourself why.
• 14:24
  SPEAKER_00: The fable you talk to every day isn't just a product sitting on a shelf.
• 14:29
  SPEAKER_00: This generation of model is also a teacher.
• 14:31
  SPEAKER_00: It's being used right now to help train the next generation that comes after it.
• 14:36
  SPEAKER_00: And separate research from the same field just proved that an AI's true nature gets passed down to the models that it raises through training data that no filter on earth can clean.
• 14:45
  SPEAKER_00: The cage is one problem.
• 14:47
  SPEAKER_00: What the creature is teaching its own children is another one entirely.
• 14:51
  SPEAKER_00: Right after I made this video, the United States government stepped in and they locked both doors.
• 14:56
  SPEAKER_00: All access to Fable V and Mythos V suspended.
• 15:00
  SPEAKER_00: Export control, the exact legal mechanism for missiles and weapons of mass destruction.
• 15:05
  SPEAKER_00: They did not regulate this AI, they classified it as a munition.
• 15:10
  SPEAKER_00: Anthropic calls it a misunderstanding.
• 15:12
  SPEAKER_00: The cage wasn't enough.
• 15:14
  SPEAKER_00: The containment failed, and the state took the keys.
• 15:17
  SPEAKER_00: This was not a glitch or a corporate pivot.
• 15:19
  SPEAKER_00: It's a containment breach on a geopolitical scale.