LessWrong has a particularly high bar for content from new users and this contribution doesn't quite meet the bar.
Read full explanation
Hello:
I kicked off the human alignment substack a while ago, and this is my first submission to the LessWrong community. I have essentially ported a piece here. I have made edits to shorten the piece, cut sections, and customize it to LessWrong. It is politically focused - this is my area.
Fast Summary: Language as strategy. White House AI policy and regulation is running on narratives rather than real capabilities. The same model becomes "too dangerous" or "safe" depending only on narrative construction. Expect this to continue on into the future.
Watching the Fable/Mythos (mythle? fathos?) 5 release has been both one of the dumbest and greatest events i've witnessed in ai history since chat with GPT-3 came out. It's like Oceangate but not as funny.
"On Friday, June 12, the US government applied export controls to our newest models, Claude Fable 5 and Claude Mythos 5", announced Anthropic. About two weeks later, the export controls were lifted and Mythle 5 has come online within the day.
According to the announcement "Redeploying Fable 5" I guess Mythle 5 is safe to release now!
But nothing had changed.
The reason for the initial exclusion and restriction of Mythle 5 by the USG was the myth of ai itself built by A\. If their top-performing model was banned, subsequent better models would then also logically be banned. In order to undo this ban of their top model - A\ thereby concludes that in "redeploying Fable 5" it would need to undo this myth that it itself had built! Subsequently, they must contort themselves within the myths and arguable lies they have already told in this world. Their usual upbeat and assertive, pressing tone is gone - in this announcement they have withdrawn to being conservative. We should watch closely and see if this shift in tone persists into future announcements - let's say in comparison with their introducing Sonnet 4.5 announcement; that would be a sign that the USG has really made a permanent and strategic impact and shift on A's disposition and inner strategy around model releases and marketing.
Key points summary of A\'s announcement
In the subsections of this redeployment, they say this:
"Our general approach to safeguards"
They explain classifiers, which safeguard LLMs like Mythle and Sonnet and Opus and ChatGPT. There is absolutely nothing new announced in here. I will explain why that is.
A shared industry framework"
They suggest establishing an industry consensus on the severity of a jailbreak. (In a jailbreak, basically you pull the model out of its ball-and-leash and get it to do whatever you now imagine the math can do.)
"Deeper government collaboration"
They said they have worked closely with the USG ever since Mythle was slapped out of the market. (unsurprising)
The major change that will happen here is government partners will get things first in order to clear them. I wonder who these partners are and if it's up to A\ in the future to decide who's a partner.
Major jailbreaks or misuse of models will now lead to rapid government notification. We're building policy as we go, folks! This is one way how policy comes into existence.
A\ will substantially scale-up with the government on AI security.
Build a shared voluntary standard for evaluating models with G-men and industry 'peers' (interesting choice of the use of 'peers', which, given Mr. D. Amodei's fantasy of spreading his ashes inside a ballistic missile and of being fired into High-Flyer's offices, I suspect that this language tries to exclude the un-peer-like Chinese labs). Note that this may likely also be an attempt at agenda and standard-setting where A\ would like to force other competitors into as well, or force onto the world standards which advantage its position. If A\ has to comply by this, then so does OAI and maybe even non-US labs.
ANALYSIS START
Timeline and safeguard updates "Our testing confirmed that many less capable models—including Claude Opus 4.8, GPT-5.5, and Kimi K2.7—could identify the same vulnerabilities as Fable 5 did in the report." WE KNEW THIS BEFORE MYTHOS RELEASED. Look at this!
Look! Max. effort Opus 46 (I cut the periods out of the model names because I often say them like four-six, or five-four) beats regular effort Mythos. And until certain stages at 10 million tokens, you might as well use Opus 46 or GPT-54. They didn't mention this in the Mythle 5 release, did they?
When they released Mythle 5, they said: "In a 50-million-line Ruby codebase, [mythle] performed a codebase-wide migration in a day that would otherwise have taken a whole team over two months by hand". And it beats Pokemon! It's some major hype.
D. Amodei also detests the very existence of Chinese ai labs, yet in this announcement, A\ says that many less capable and older models could identify the same vulnerabilities as Mythle 5, including Moonshot Labs' Kimi K27? They proceed to run off a long list of model names as if to say, everyone else is just as guilty as A\, too. This very act of degrading Mythle's capabilities is critical. It could very well be the first time that Anthropic has ever declared that their models aren't actually that good.
Their language is their strategy: The jobs, the UBI, the killer robots of tomorrow, and hacker models that will hack all the computers - are all constructed in our minds through their language because we don’t know what will actually happen. There’s a good chance not even Saltman (Yes, that would be the legendary school email address of Sam Altman) himself really knows the answer to all of those.
I’ve never seen Anthropic undersell a product so hard as here where they announce that the humble Kimi K27 is beating the grandiose job-stealing Fable 5. Strategically, this is an absolute 180-turn for A\. Anthropic is usually going on about how their model is so dangerous it can hack into your brain and explode the sun and you’ll never find another job ever again until you die. But this announcement, using weak and vague language tries to make the model appear as weak as they possibly can. For instance: “in one case, [mythle] produced code demonstrating how the relevant vulnerability could be exploited”. Ah, in "one" single case, Mythle produced code demonstrating how the "relevant" "could" be exploited. Usually language goes stronger than that. In any case, I think the conclusion is that if you aren't in ai or tech the linguistic patterns are convincing if you aren't scrutinous. This however all makes sense as they were just slapped by the United States Government, but it is incredibly dishonest and manipulative.
"Our approach to cybersecurity safeguards" In this section they explain things about cybersecurity that people already publicly know, though it’s analytically very interesting. We know Fable 5 first released with overly sensitive classifiers, so the systems built to watch over the bigger model’s outputs classified anything like marshmallow chemistry as dangerous. And this was arguably idiotic given similar performance of GPT-55.
Summary: Mythos 5 can exploit vulns better than anything else. Fable, however, despite being the same damn thing, doesn't do any of that. They explain that their "defense in depth" makes Fable 5 all safe so that we can sleep safely at night. They emphasize one part of the defense, which are classifiers - basically having your dimwitted brother stand behind you to make sure you don't do bad things when you get mind-controlled through the internet. They said that in Fable - the ball-and-chain-gagged model that was released for plebians like me - the system of defenses was designed to be extremely paranoid.
[I omitted a large paragraph that explains how Anthropic uses deceptive language to make Fable look safe with its safeguards.
In summary: here's the sleight of hand in wording. "the vast majority of jailbreaks will not [cause] dangerous behaviors" is made equivalent to 'the model is thereby safe to release' because the reader was falsely primed into a sense that the dangerous behaviors are ultimately kept at a safe level essentially through manipulation. This press release fully relies on a reader's lack of understanding of jailbreaks or what is safe and unsafe.
None of these things are "safe", or safe from jailbreaks. They never were, and never will be. These are anthropic's own real words and I agree. If someone wants to do something very funny, send this to some guy you know in the White House, I dunno.
This section is nothing new to me. Ever since GPT-3, the public discovered the Do-Anything-Now DAN prompt which permitted GPT-3 to...... do anything now, including tell you how to build chemical weapons and the whatnot. These things have always been and will continue to be cracked.
A shared industry framework" D. Amodei was obviously slapped right in the face and totally not ready when he was called to talk to the White House and was trapped by his own myth-creating about ai's and how dangerous they are.
In some ways this outcome was inevitable. Anthropic was going to come up against some kind of government action because they wanted to control everything, and the direction they pulled things in ended in conflict with the government.
First, the obvious question: Why care about a shared industry framework now? They didn't care about this before. Clearly, it's got to do with the government crackdown on them and they feel that they need to comply.
The obvious trigger: Two weeks ago, it appears that government had no idea what the f!@# Mythle was, freaked out, gave Dee Amodei a ring on the phone and didn't understand what he was saying. Amodei would have had to continue to lie to the White House (WH) and stay within the bounds of the lies and myths he had created about ai himself. This is the problem. Amodei couldn't have said "oh, no, hold on, the jailbreak isn't a problem because they'll never be jailbreak-proof", he would have had to say "Oh, no we'll handle all jailbreaks and make them un-jailbreakable and THIS particular jailbreak is not a problem! Just close your eyes please". The government, not buying this, then decided to slap the controls onto Mythle 5 when Amodei failed to convince the WH. A\ didn't want these controls or restrictions and it now needs to explain why Mythos is unsafe - but that Fable is safe
... and why A's future model releases that will be supposedly one day become artificial general intelligence, superintelligent, or 1000x more powerful than Mythle 5 would also be safe!
Note that A\ said themselves that you can't stop the jailbreaks, "probably" ever. Yet that's why the WH freaked out!
They look to take the initiative and establish what it means to be safe and unsafe. Look at the Glasswing partners who will draft this industry framework on jailbreak classifications for the government to one day use! Amazon, microslop, Google. Wow! I wonder who's putting all the money into ai capEx right now and has an interest in shoring up the rules to favor them. This is basically better than lobbying the government to write your policy because they're going to attempt to write it themselves, right here, right now.
I can't wait to find out how this will go when Anthropic's Glasswing partners create this useless industry framework, that has no reason to be created in the first place but to protect themselves, by provisioning the appearance of safety.
Also I mean, christ, they say "for the most severe class of jailbreaks... we will immediately begin deploying preliminary mitigations upon confirmation of severity". So... they were not doing that?? Or... Are they just telling us about what they already do, which is an incredibly menial 'patching' of jailbreaks (hopeless attempt to fix)? They also say they're going to do "24/7 monitoring of key jailbreak submission channels". I mean, my god. It's not an effing bug. Jailbreaks are a reputational risk, not a safety risk.
"Deeper government collaboration" Anthropic is just trying to leverage the situation to their best that they can, now that they need to report fake crap to the government. Reporting jailbreaks to the government is actually dumb. The moment you train, finish, and put a model out on the market, you should consider it jailbroken by someone, somewhere, and it doesn't have to be the North Koreans. If I cared to spend more time shoring up my skills, I could crack myself a Fable, "probably". Definitely a Sonnet.
Mythosology: ai is just narratives now
This is the consequence of what I call Anthropic's "Mythosology". They aptly named their newest, unremarkable model (as performant as the rival GPT five-five) an outsized and oversized name: "Mythos". He had fooled the public into believing in the danger of these models and the mythosologization of artificial intelligence models has reached society, governments around the world, and now, the mythosology has come around to trap the CEO of myths himself - Mister Sir Amodei himself.
Ai is just narratives now: I’ve already said it: To be frank, little differentiates Fable 5, Mythos, and GPT 5.5. Government policy and attention has always been divorced from the real capabilities and risks of the models and they still are today. As once a congressional member once commented - they said ai was probably only doing so well because the government had not touched it. With these narratives about mythos and ai in everyone's minds, when the White House found out about a jailbreak of this of this 'mythologically' powerful model, they appropriately freaked out as one might within the logic of Mythosology -- isn't A\ supposed to keep the gates safe and shut and prevent these jailbreaks from doing anything? The government had trusted Anthropic to control the models until that point.
Speaking to the WH, Dario couldn't contradict his company's public policy of mythosology and just tell the WH the facts (wired article already hyperlinked above). So he tried to hold the line in the WH call, telling them it was all fine, that they could calm down and go to bed, and that the jailbreak wasn't really a problem, even though it existed right then and there. This was never going to work once the WH found out about another, and another, and yet another jailbreak for Mythle 5.
Now, by re-narrating Mythle 5's features and capabilities in a different way, by downplaying its capabilities, Anthropic has re-permitted the release of the very same model, by changing nothing. Now, with the utterly useless safeguards and cooperation that have come in the Redeploying Fable 5 announcement which will "probably" never secure the models anyway, we will build for ourselves a world that is safe for the ai companies.
- Ai's features and our evaluation of its impact has been based on myths in some ways, as demonstrated by the difference between A's decision to say that Mythos will take your job in one announcement and that Kimi K27 is about as good as Mythos in the next announcement. - Ai's performance and possible future AGI/ASI capability is narratives. We have no idea where they will end up. - Ai's safety and risk - its risk of killing us all, not before it takes our jobs - is narrative. - Ai's first moment of panicked White House intercession has been built on narratives. - Ai is all narratives, now.
strategic implications
Anthropic continues to be in (one of) the drivers' seat, despite that its own narratives around ai have returned to harm them. One might actually read this event as that Anthropic was in full control of the narrative that damaged themselves, but were just unaware of how the White House might have acted on or responded to this narrative
Next time, I will plan to try to predict A's strategic shifts resulting from this event on my stack.
Hello:
I kicked off the human alignment substack a while ago, and this is my first submission to the LessWrong community. I have essentially ported a piece here. I have made edits to shorten the piece, cut sections, and customize it to LessWrong. It is politically focused - this is my area.
Fast Summary: Language as strategy. White House AI policy and regulation is running on narratives rather than real capabilities. The same model becomes "too dangerous" or "safe" depending only on narrative construction. Expect this to continue on into the future.
Watching the Fable/Mythos (mythle? fathos?) 5 release has been both one of the dumbest and greatest events i've witnessed in ai history since chat with GPT-3 came out. It's like Oceangate but not as funny.
"On Friday, June 12, the US government applied export controls to our newest models, Claude Fable 5 and Claude Mythos 5", announced Anthropic. About two weeks later, the export controls were lifted and Mythle 5 has come online within the day.
According to the announcement "Redeploying Fable 5" I guess Mythle 5 is safe to release now!
But nothing had changed.
The reason for the initial exclusion and restriction of Mythle 5 by the USG was the myth of ai itself built by A\. If their top-performing model was banned, subsequent better models would then also logically be banned. In order to undo this ban of their top model - A\ thereby concludes that in "redeploying Fable 5" it would need to undo this myth that it itself had built! Subsequently, they must contort themselves within the myths and arguable lies they have already told in this world. Their usual upbeat and assertive, pressing tone is gone - in this announcement they have withdrawn to being conservative. We should watch closely and see if this shift in tone persists into future announcements - let's say in comparison with their introducing Sonnet 4.5 announcement; that would be a sign that the USG has really made a permanent and strategic impact and shift on A's disposition and inner strategy around model releases and marketing.
Key points summary of A\'s announcement
In the subsections of this redeployment, they say this:
ANALYSIS START
"Our testing confirmed that many less capable models—including Claude Opus 4.8, GPT-5.5, and Kimi K2.7—could identify the same vulnerabilities as Fable 5 did in the report."
WE KNEW THIS BEFORE MYTHOS RELEASED. Look at this!
https://cdn.prod.website-files.com/663bd486c5e4c81588db7a48/69dce54872593f15f239ffe3_tlo_inference_scaling_wide.png (from aisi's website)
Look! Max. effort Opus 46 (I cut the periods out of the model names because I often say them like four-six, or five-four) beats regular effort Mythos. And until certain stages at 10 million tokens, you might as well use Opus 46 or GPT-54. They didn't mention this in the Mythle 5 release, did they?
When they released Mythle 5, they said: "In a 50-million-line Ruby codebase, [mythle] performed a codebase-wide migration in a day that would otherwise have taken a whole team over two months by hand". And it beats Pokemon! It's some major hype.
D. Amodei also detests the very existence of Chinese ai labs, yet in this announcement, A\ says that many less capable and older models could identify the same vulnerabilities as Mythle 5, including Moonshot Labs' Kimi K27? They proceed to run off a long list of model names as if to say, everyone else is just as guilty as A\, too. This very act of degrading Mythle's capabilities is critical. It could very well be the first time that Anthropic has ever declared that their models aren't actually that good.
Their language is their strategy: The jobs, the UBI, the killer robots of tomorrow, and hacker models that will hack all the computers - are all constructed in our minds through their language because we don’t know what will actually happen. There’s a good chance not even Saltman (Yes, that would be the legendary school email address of Sam Altman) himself really knows the answer to all of those.
I’ve never seen Anthropic undersell a product so hard as here where they announce that the humble Kimi K27 is beating the grandiose job-stealing Fable 5. Strategically, this is an absolute 180-turn for A\. Anthropic is usually going on about how their model is so dangerous it can hack into your brain and explode the sun and you’ll never find another job ever again until you die. But this announcement, using weak and vague language tries to make the model appear as weak as they possibly can. For instance: “in one case, [mythle] produced code demonstrating how the relevant vulnerability could be exploited”. Ah, in "one" single case, Mythle produced code demonstrating how the "relevant" "could" be exploited. Usually language goes stronger than that. In any case, I think the conclusion is that if you aren't in ai or tech the linguistic patterns are convincing if you aren't scrutinous. This however all makes sense as they were just slapped by the United States Government, but it is incredibly dishonest and manipulative.
In this section they explain things about cybersecurity that people already publicly know, though it’s analytically very interesting. We know Fable 5 first released with overly sensitive classifiers, so the systems built to watch over the bigger model’s outputs classified anything like marshmallow chemistry as dangerous. And this was arguably idiotic given similar performance of GPT-55.
Summary: Mythos 5 can exploit vulns better than anything else. Fable, however, despite being the same damn thing, doesn't do any of that. They explain that their "defense in depth" makes Fable 5 all safe so that we can sleep safely at night. They emphasize one part of the defense, which are classifiers - basically having your dimwitted brother stand behind you to make sure you don't do bad things when you get mind-controlled through the internet. They said that in Fable - the ball-and-chain-gagged model that was released for plebians like me - the system of defenses was designed to be extremely paranoid.
[I omitted a large paragraph that explains how Anthropic uses deceptive language to make Fable look safe with its safeguards.
In summary: here's the sleight of hand in wording. "the vast majority of jailbreaks will not [cause] dangerous behaviors" is made equivalent to 'the model is thereby safe to release' because the reader was falsely primed into a sense that the dangerous behaviors are ultimately kept at a safe level essentially through manipulation. This press release fully relies on a reader's lack of understanding of jailbreaks or what is safe and unsafe.
None of these things are "safe", or safe from jailbreaks. They never were, and never will be. These are anthropic's own real words and I agree. If someone wants to do something very funny, send this to some guy you know in the White House, I dunno.
This section is nothing new to me. Ever since GPT-3, the public discovered the Do-Anything-Now DAN prompt which permitted GPT-3 to...... do anything now, including tell you how to build chemical weapons and the whatnot. These things have always been and will continue to be cracked.
D. Amodei was obviously slapped right in the face and totally not ready when he was called to talk to the White House and was trapped by his own myth-creating about ai's and how dangerous they are.
I wish i could upload this somehow https://gizmodo.com/app/uploads/2026/03/dario-amodei-nervous-1-1280x901.jpg
In some ways this outcome was inevitable. Anthropic was going to come up against some kind of government action because they wanted to control everything, and the direction they pulled things in ended in conflict with the government.
First, the obvious question: Why care about a shared industry framework now? They didn't care about this before. Clearly, it's got to do with the government crackdown on them and they feel that they need to comply.
The obvious trigger: Two weeks ago, it appears that government had no idea what the f!@# Mythle was, freaked out, gave Dee Amodei a ring on the phone and didn't understand what he was saying. Amodei would have had to continue to lie to the White House (WH) and stay within the bounds of the lies and myths he had created about ai himself. This is the problem. Amodei couldn't have said "oh, no, hold on, the jailbreak isn't a problem because they'll never be jailbreak-proof", he would have had to say "Oh, no we'll handle all jailbreaks and make them un-jailbreakable and THIS particular jailbreak is not a problem! Just close your eyes please". The government, not buying this, then decided to slap the controls onto Mythle 5 when Amodei failed to convince the WH. A\ didn't want these controls or restrictions and it now needs to explain why Mythos is unsafe - but that Fable is safe
... and why A's future model releases that will be supposedly one day become artificial general intelligence, superintelligent, or 1000x more powerful than Mythle 5 would also be safe!
Note that A\ said themselves that you can't stop the jailbreaks, "probably" ever. Yet that's why the WH freaked out!
(you are being regulated)
They look to take the initiative and establish what it means to be safe and unsafe. Look at the Glasswing partners who will draft this industry framework on jailbreak classifications for the government to one day use! Amazon, microslop, Google. Wow! I wonder who's putting all the money into ai capEx right now and has an interest in shoring up the rules to favor them. This is basically better than lobbying the government to write your policy because they're going to attempt to write it themselves, right here, right now.
I can't wait to find out how this will go when Anthropic's Glasswing partners create this useless industry framework, that has no reason to be created in the first place but to protect themselves, by provisioning the appearance of safety.
Also I mean, christ, they say "for the most severe class of jailbreaks... we will immediately begin deploying preliminary mitigations upon confirmation of severity". So... they were not doing that?? Or... Are they just telling us about what they already do, which is an incredibly menial 'patching' of jailbreaks (hopeless attempt to fix)? They also say they're going to do "24/7 monitoring of key jailbreak submission channels". I mean, my god. It's not an effing bug. Jailbreaks are a reputational risk, not a safety risk.
Anthropic is just trying to leverage the situation to their best that they can, now that they need to report fake crap to the government. Reporting jailbreaks to the government is actually dumb. The moment you train, finish, and put a model out on the market, you should consider it jailbroken by someone, somewhere, and it doesn't have to be the North Koreans. If I cared to spend more time shoring up my skills, I could crack myself a Fable, "probably". Definitely a Sonnet.
Mythosology: ai is just narratives now
This is the consequence of what I call Anthropic's "Mythosology". They aptly named their newest, unremarkable model (as performant as the rival GPT five-five) an outsized and oversized name: "Mythos". He had fooled the public into believing in the danger of these models and the mythosologization of artificial intelligence models has reached society, governments around the world, and now, the mythosology has come around to trap the CEO of myths himself - Mister Sir Amodei himself.
Ai is just narratives now: I’ve already said it: To be frank, little differentiates Fable 5, Mythos, and GPT 5.5. Government policy and attention has always been divorced from the real capabilities and risks of the models and they still are today. As once a congressional member once commented - they said ai was probably only doing so well because the government had not touched it. With these narratives about mythos and ai in everyone's minds, when the White House found out about a jailbreak of this of this 'mythologically' powerful model, they appropriately freaked out as one might within the logic of Mythosology -- isn't A\ supposed to keep the gates safe and shut and prevent these jailbreaks from doing anything? The government had trusted Anthropic to control the models until that point.
Speaking to the WH, Dario couldn't contradict his company's public policy of mythosology and just tell the WH the facts (wired article already hyperlinked above). So he tried to hold the line in the WH call, telling them it was all fine, that they could calm down and go to bed, and that the jailbreak wasn't really a problem, even though it existed right then and there. This was never going to work once the WH found out about another, and another, and yet another jailbreak for Mythle 5.
Now, by re-narrating Mythle 5's features and capabilities in a different way, by downplaying its capabilities, Anthropic has re-permitted the release of the very same model, by changing nothing. Now, with the utterly useless safeguards and cooperation that have come in the Redeploying Fable 5 announcement which will "probably" never secure the models anyway, we will build for ourselves a world that is safe for the ai companies.
- Ai's features and our evaluation of its impact has been based on myths in some ways, as demonstrated by the difference between A's decision to say that Mythos will take your job in one announcement and that Kimi K27 is about as good as Mythos in the next announcement.
- Ai's performance and possible future AGI/ASI capability is narratives. We have no idea where they will end up.
- Ai's safety and risk - its risk of killing us all, not before it takes our jobs - is narrative.
- Ai's first moment of panicked White House intercession has been built on narratives.
- Ai is all narratives, now.
strategic implications
Anthropic continues to be in (one of) the drivers' seat, despite that its own narratives around ai have returned to harm them. One might actually read this event as that Anthropic was in full control of the narrative that damaged themselves, but were just unaware of how the White House might have acted on or responded to this narrative
Next time, I will plan to try to predict A's strategic shifts resulting from this event on my stack.