5 ms·

I am really confused as to what happened here. The use of ‘disabled organization’ to refer to the author made it extra confusing. I think I kind of have an ide

I am really confused as to what happened here. The use of ‘disabled organization’ to refer to the author made it extra confusing. I think I kind of have an idea what the author was doing, but not really.

tobyhinloopen 9mo ago
I had to read it twice as well, I was so confused hah. I’m still confused
- rtkwe 9mo ago
  They probably organize individual accounts the same as organization accounts for larger groups of users at the same company internally since it all rolls up to one billing. That's my first pass guess at least.
alistairSH 9mo ago
You're not alone. I think the author was doing some sort of circular prompt injection between two instances of Claude? The author claims "I'm just scaffolding a project" but that doesn't appear to be the case, or what resulted in the ban...
- lazyfanatic42 9mo ago
  Author really comes off unhinged throughout the article to be frank.
  - superb_dev 9mo ago
    Did we read the same article? The author comes of as pretty frustrated but not unhinged
    - ryandrake 9mo ago
      I wouldn't say "unhinged" either, but maybe just struggling to organize and express thoughts clearly in writing. "Organizations of late capitalism, unite"?
      - Bootvis 9mo ago
        The author was frustrated that the error message identified him as an organisation (that was disabled) and mockingly refers to himself as the (disabled) organisation in the post. At least, that’s my reading but it appears it confuses about half of the commenters here.
        ryandrake 9mo ago
        I think if one's readers need an "ironic euphemism decoder glossary" just to understand the message, it could use a little re-writing.
        layer8 9mo ago
        It was perfectly understandable to me. Maybe cultural differences? You seem to be American, OP Portuguese, and myself European as well.
        ashirviskas 9mo ago
        Another European chiming in, I enjoyed OPs article.
        superb_dev 9mo ago
        I’m American and it made sense
      - genewitch 9mo ago
        https://en.wikipedia.org/wiki/Late_capitalism https://en.wikipedia.org/wiki/Late_capitalism https://community.bitwarden.com/t/re-enabling-a-disabled-organization/67466 https://community.bitwarden.com/t/re-enabling-a-disabled-org... https://community.meraki.com/t5/Dashboard-Administration/disabled-organization/m-p/269725 https://community.meraki.com/t5/Dashboard-Administration/dis... the former i have heard for a couple decades, the latter is apparently a term of art to prevent hurt feelings or lawsuits or something. Google thinks i want ADA style organizations, but it's AI caught on that i might not mean organizations for disabled people btw "ADA" means Americans with Disabilities Act. AI means Artificial Intelligence. A decade is 10 years long. "term of art" is a term of art for describing stuff like jargon or lingo of a trade, skill, profession. Jargon is specialized, technical language used in a field or area of study. Lingo pins to jargon, but is less technical. Google is a company that started out crawling the web and making a web search site that they called a search engine. They are now called Alphabet Company (ABC). Crawling means to iteratively parse the characters sent by a webserver and follow links therein, keeping a copy of the text from each such html. HTML is hypertext markup language, hypertext is like text, but more so. Language is how we communicate. I can go on? p.s. if you want a better word, your complaint is about the framing. you didn't gel with the framing of the article. My friend, who holds a doctorate, defended a thesis about how virtually every platform argument is really a framing issue. platform as in, well, anything you care to defend. mac vs linux, wifi vs ethernet, podcasts vs music, guns vs no guns, red vs blue. If you can reduce the frame of the context to something both parties can agree to, you can actually hold a real, intellectual debate, and get at real issues.
  - pjbeam 9mo ago
    My take was more a kind of amusing laughing-through-frustration but also enjoying the ride just a little bit insouciance. Tastes vary of course, but I enjoyed the author's tone and pacing.
  - staticman2 9mo ago
    Author thinks he's cute to do things like mention Google without typing Google but I wouldn't call him unhinged.
- rvba 9mo ago
  What is wrong with circular prompt injection? The "disabled organization" looks like a sarcastic comment on the crappy error code the author got when banned.
  - darkwater 9mo ago
    > What is wrong with circular prompt injection? That you might be trying to jailbreak Claude and Anthropic does not like that (I'm not endorsing, just trying to understand).
- redeeman 9mo ago
  i have no idea what he was actually doing either, and what exactly is it one isnt allowed to use claude to do?
- falloutx 9mo ago
  This tracks with Anthropic, they are actively hostile to security researchers.
- Romario77 9mo ago
  One Claude agent told other Claude agent via CLAUDE.md to do things certain way. The way Claude did it triggered the ban - i.e. it used all caps which apparently triggers some kind of internal alert, Anthropic probably has some safeguards to prevent hacking/prompt injection and what the first Claude did to CLAUDE.md triggered this safeguard. And it doesn't look like it was a proper use of the safeguard, they banned for no good reason.
- layer8 9mo ago
  It wasn’t circular. TFA explains how the author was always in the loop. He had one Claude instance rewrite the CLAUDE.MD of another Claude instance whenever the second one made a mistake, but relaying the mistake to the first instance (after recognizing it in the first place) was done manually by the author.
- healsdata 9mo ago
  The author code have easily shared the last version of Claude.md that had the all caps or whatever, but didn't. Points to something fishy in my mind.
  - BoorishBears 9mo ago
    The whole thing reads like LLM psychosis.
  - ribosometronome 9mo ago
    They did. >If you want to take a look at the CLAUDE.md that Claude A was making Claude B run with, I commited it and it is available here. https://github.com/HugoDaniel/boreDOM/blob/9a0802af16f5a1ff177799404c34ce5444345915/boreDOMCLI/cli.js#L129 https://github.com/HugoDaniel/boreDOM/blob/9a0802af16f5a1ff1...
- cryptonector 9mo ago
  I suspeect that having Claudes talking to Claudes is a very bad idea from Anthropic's point of view because that could easily consume a ton of resources doing nothing useful.
anigbrowl 9mo ago
Agreed, I found this rather incoherent and seeming to depend on knowing a lot more about author's project/background.
cr3ative 9mo ago
Right. This is almost unreadable. There are words, but the author seems to be too far down a rabbit hole to communicate the problem properly…
superb_dev 9mo ago
The author was using instance A of Claude to update a `claude.md` while another instance B of Claude was consuming that file. When Claude B did something wrong, the author asked Claude A to update the `claude.md` so that Claude B didn’t make the same mistake again
- raincole 9mo ago
  Which shouldn't be bannable imo. Rate throttle is a more reasonable response. But Anthropic didn't reply to the author, so we don't even know if it's the real reason they got banned.
  - pocksuppet 9mo ago
    When a company won't tell you what you did wrong, you should be free to take the least charitable interpretation towards the company. If it was more charitable, they'd tell you.
  - pixl97 9mo ago
    >if it's the real reason they got banned. I mean, what a country should do it put a law in effect. If you ban a user, the user can submit a request with their government issued ID and you must give an exact reason why they were banned. The company can keep this record in encrypted form for 10 years. Failure to give the exact reason will lead to a $100,000 fine for the first offense and increase from there up to suspension of operations privileges in said country. "But, but, but hackers/spammers will abuse this". For one, boo fucking hoo. For two, just add to the bill "Fraudulent use of law to bypass system restrictions is a criminal offense". This puts companies in a position where they must be able to justify their actual actions, and it also puts scammers at risk if they abuse the system.
    - benjiro 9mo ago
      Companies will simply give some kind of standard answer, that is legally "cover our butts" and be done with it. Its like that cookie wall stuff, how much dark patterns are implemented. They followed the letter of the law, not the spirit of the law. To be honest, i can also see the point from the company side. Giving a honest answer can just anger people, to the point they sue. People are often not as rational as we all like our fellow humans to be. Even if the ex-client lose in court, that is how much time you wasted on issue clients... Its one thing if your a big corporation with tons of lawyers but small companies are often not in the position to deal with that drama. And it can take years to resolve. Every letter, every phone call to a lawyer, it stacks up fast! Do you get your money back? Maybe, depends on the country, but your time? I am not pro companies but its often simply better to have the attitude "you do not want me as your client, let me advocate for your competitor and go there".
exitb 9mo ago
Normally you can customize the agents behavior via a CLAUDE.md file. OP automated that process by having another agent customize the first agent. The customizer agent got pushy, the customized agent got offended, OP got banned.
Aurornis 9mo ago
Years ago I was involved in a service where we some times had to disable accounts for abusive behavior. I'm talking about obvious abusive behavior, akin to griefing other users. Every once in while someone would take it personally and go on a social media rampage. The one thing I learned from being on the other side of this is that if someone seems like an unreliable narrator, they probably are. They know the company can't or won't reveal the true reason they were banned, so they're virtually free to tell any story they want. There are so many things about this article that don't make sense: > I'm glad this happened with this particular non-disabled-organization. Because if this by chance had happened with the other non-disabled-organization that also provides such tools... then I would be out of e-mail, photos, documents, and phone OS. I can't even understand what they're trying to communicate. I guess they're referring to Google? There is, without a doubt, more to this story than is being relayed.
- dragonwriter 9mo ago
  The excerpt you don’t understand is saying that if it has been Google rather than Anthropic, the blast radius of the no-explanation account nuking would have been much greater. It’s written deliberately elliptically for humorous effect (which, sure, will probably fall flat for a lot of people), but the reference is unmistakable.
- fluoridation 9mo ago
  "I'm glad this happened with Anthropic instead of Google, which provides Gemini, email, etc. or I would have been locked out of the actually important non-AI services as well." Non-disabled organization = the first party provider Disabled organization = me I don't know why they're using these weird euphemisms or ironic monikers, but that's what they mean.
  - mattnewton 9mo ago
    Because they bought a claude subscription on a personal account and the error message said that they belongs to a "disabled organization" (probably leaking some implementation details).
Romario77 9mo ago
You are confused because the message from Claude is confusing. Author is not an organization, they had an account with anthropic which got disabled and Anthropic addressed them as organization.
- dragonwriter 9mo ago
  > Author is not an organization, they had an account with anthropic which got disabled and Anthropic addressed them as organization. Anthropic accounts are always associated with an organization; for personal accounts the Organization and User name are identical. If you have an Anthropic API account, you can verify this in the Settings pane of the Dashboard (or even just look at the profile button which shows the org and account name.)
  - ryandrake 9mo ago
    I've always kind of hated that anti-pattern in other software I use for peronal/hobby purposes, too. "What is your company name? [required]" I don't have a company! I'm just playing around with your tool on my own! I'm not an organization!
ankit219 9mo ago
My rudimentary guess is this. When you write in all caps, it triggers sort of a alert at Anthropic, especially as an attempt to hijack system prompt. When one claude was writing to other, it resorted to all caps, which triggered the alert, and then the context was instructing the model to do something (which likely would be similar to a prompt injection attack) and that triggered the ban. not just caps part, but that in combination of trying to change the system characteristics of claude. OP does not know much better because it seems he wasn't closely watching what claude was writing to other file. if this is true, the learning is opus 4.5 can hijack system prompts of other models.
- kstenerud 9mo ago
  > When you write in all caps, it triggers sort of a alert at Anthropic I find this confusing. Why would writing in all caps trigger an alert? What danger does caps incur? Does writing in caps make a prompt injection more likely to succeed?
  - ankit219 9mo ago
    from what i know, it used to be that if you want to assertively instruct, you used all caps. I don't know if it succeeds today. I still see prompts where certain words are capitalized to ensure model pays attention. What i mean was not just capitalization, but a combination of both capitalization and changing the behavior of the model for trying to get it to do something. if you were to design a system to prevent prompt injections and one of surefire ways is to repeatedly give instructions in caps, you would have systems dealing with it. And with instructions to change behavior, it cascades.
  - direwolf20 9mo ago
    Many jailbreaks use allcaps
- phreack 9mo ago
  Wait what? Really? All caps is a bannable offense? That should be in all caps, pardon me, in the terms of use if that's the case. Even more so since there's no support at the highest price point.
mmkos 9mo ago
You and me, brother. The writing is unnecessarily convoluted.
vimda 9mo ago
Yeah, referring to yourself once as a "disabled organisation" is a good bit, referencing anthropics silly terminology. Keeping it for the duration made this a very hard follow
- Ronsenshi 9mo ago
  Sounds like author of the post might have needed an AI to review and fix his convoluted writing. Maybe even two AIs!
- llIIllIIllIIl 9mo ago
  On the contrary I enjoyed this human touch in the text.
verdverm 9mo ago
Sounds like OP has multiple org accounts with Anthropic. The main one in the story (disabled) is banned because iterating on claude.md files looks a lot like iterating on prompt injections, especially as it sounds the multiple Claude's got into it with each other a bit The other org sounds like the primary account with all the important stuff. Good on OP for doing this work in a separate org, a good recommendation across a lot of vendors and products.
alasr 9mo ago
> I think I kind of have an idea what the author was doing, but not really. Me neither; However, just like the rest I can only speculate (given the available information): I guess the following pieces provide a hint what's really going on here: - "The quine is the quine" (one of the sub-headline of the article) and the meaning of the word "quine". - Author's "scaffolding" tool which, once finished, had acquired the "knowledge"[1] how to add a CLAUDE.md baked instructions for a particular homemade framework (he's working on). - Anthropic saying something like: no, stop; you cannot "copy"[1] Claude knowledge no matter how "non-serious" your scaffolding tool or your use-case is: as it might "shows", other Claude users, that there's a way to do similar things, maybe that time, for more "serious" tools. --- [1]. Excerpt from the Author's blog post: "I would love to see the face of that AI (Claude AI system backend) when it saw its own 'system prompt' language being echoed back to it (from Author's scaffolding tool: assuming it's complete and fully-functional at that time)."
NBJack 9mo ago
I think you missed the joke: he isn't an organization at all, but the error message claims he is.
saghm 9mo ago
From reading the whole thing, it kind of seems clickbaity. Yes, they're the only user in the "organization" that got banned, but they apparently still are using the other "organization" without issue, so they as a human are not banned. There's certainly a valid complaint to be made about the lack of recourse or customer service response for the automated ban, but it almost seems like they intentionally were trying to be misleading by implying that since they were the only member of the organization, they were banned from using Claude.
aswegs8 9mo ago
Yeah, I couldn't follow this "disabled organization" and "non-disabled organization" naming either.

tobyhinloopen 9mo ago

I had to read it twice as well, I was so confused hah. I’m still confused

rtkwe 9mo ago
They probably organize individual accounts the same as organization accounts for larger groups of users at the same company internally since it all rolls up to one billing. That's my first pass guess at least.

alistairSH 9mo ago

You're not alone. I think the author was doing some sort of circular prompt injection between two instances of Claude? The author claims "I'm just scaffolding a project" but that doesn't appear to be the case, or what resulted in the ban...

lazyfanatic42 9mo ago
Author really comes off unhinged throughout the article to be frank.
- superb_dev 9mo ago
  Did we read the same article? The author comes of as pretty frustrated but not unhinged
  - ryandrake 9mo ago
    I wouldn't say "unhinged" either, but maybe just struggling to organize and express thoughts clearly in writing. "Organizations of late capitalism, unite"?
    - Bootvis 9mo ago
      The author was frustrated that the error message identified him as an organisation (that was disabled) and mockingly refers to himself as the (disabled) organisation in the post. At least, that’s my reading but it appears it confuses about half of the commenters here.
      - ryandrake 9mo ago
        I think if one's readers need an "ironic euphemism decoder glossary" just to understand the message, it could use a little re-writing.
        layer8 9mo ago
        It was perfectly understandable to me. Maybe cultural differences? You seem to be American, OP Portuguese, and myself European as well.
        ashirviskas 9mo ago
        Another European chiming in, I enjoyed OPs article.
        superb_dev 9mo ago
        I’m American and it made sense
    - genewitch 9mo ago
      https://en.wikipedia.org/wiki/Late_capitalism https://en.wikipedia.org/wiki/Late_capitalism https://community.bitwarden.com/t/re-enabling-a-disabled-organization/67466 https://community.bitwarden.com/t/re-enabling-a-disabled-org... https://community.meraki.com/t5/Dashboard-Administration/disabled-organization/m-p/269725 https://community.meraki.com/t5/Dashboard-Administration/dis... the former i have heard for a couple decades, the latter is apparently a term of art to prevent hurt feelings or lawsuits or something. Google thinks i want ADA style organizations, but it's AI caught on that i might not mean organizations for disabled people btw "ADA" means Americans with Disabilities Act. AI means Artificial Intelligence. A decade is 10 years long. "term of art" is a term of art for describing stuff like jargon or lingo of a trade, skill, profession. Jargon is specialized, technical language used in a field or area of study. Lingo pins to jargon, but is less technical. Google is a company that started out crawling the web and making a web search site that they called a search engine. They are now called Alphabet Company (ABC). Crawling means to iteratively parse the characters sent by a webserver and follow links therein, keeping a copy of the text from each such html. HTML is hypertext markup language, hypertext is like text, but more so. Language is how we communicate. I can go on? p.s. if you want a better word, your complaint is about the framing. you didn't gel with the framing of the article. My friend, who holds a doctorate, defended a thesis about how virtually every platform argument is really a framing issue. platform as in, well, anything you care to defend. mac vs linux, wifi vs ethernet, podcasts vs music, guns vs no guns, red vs blue. If you can reduce the frame of the context to something both parties can agree to, you can actually hold a real, intellectual debate, and get at real issues.
- pjbeam 9mo ago
  My take was more a kind of amusing laughing-through-frustration but also enjoying the ride just a little bit insouciance. Tastes vary of course, but I enjoyed the author's tone and pacing.
- staticman2 9mo ago
  Author thinks he's cute to do things like mention Google without typing Google but I wouldn't call him unhinged.
rvba 9mo ago
What is wrong with circular prompt injection? The "disabled organization" looks like a sarcastic comment on the crappy error code the author got when banned.
- darkwater 9mo ago
  > What is wrong with circular prompt injection? That you might be trying to jailbreak Claude and Anthropic does not like that (I'm not endorsing, just trying to understand).
redeeman 9mo ago
i have no idea what he was actually doing either, and what exactly is it one isnt allowed to use claude to do?
falloutx 9mo ago
This tracks with Anthropic, they are actively hostile to security researchers.
Romario77 9mo ago
One Claude agent told other Claude agent via CLAUDE.md to do things certain way. The way Claude did it triggered the ban - i.e. it used all caps which apparently triggers some kind of internal alert, Anthropic probably has some safeguards to prevent hacking/prompt injection and what the first Claude did to CLAUDE.md triggered this safeguard. And it doesn't look like it was a proper use of the safeguard, they banned for no good reason.
layer8 9mo ago
It wasn’t circular. TFA explains how the author was always in the loop. He had one Claude instance rewrite the CLAUDE.MD of another Claude instance whenever the second one made a mistake, but relaying the mistake to the first instance (after recognizing it in the first place) was done manually by the author.
healsdata 9mo ago
The author code have easily shared the last version of Claude.md that had the all caps or whatever, but didn't. Points to something fishy in my mind.
- BoorishBears 9mo ago
  The whole thing reads like LLM psychosis.
- ribosometronome 9mo ago
  They did. >If you want to take a look at the CLAUDE.md that Claude A was making Claude B run with, I commited it and it is available here. https://github.com/HugoDaniel/boreDOM/blob/9a0802af16f5a1ff177799404c34ce5444345915/boreDOMCLI/cli.js#L129 https://github.com/HugoDaniel/boreDOM/blob/9a0802af16f5a1ff1...
cryptonector 9mo ago
I suspeect that having Claudes talking to Claudes is a very bad idea from Anthropic's point of view because that could easily consume a ton of resources doing nothing useful.

anigbrowl 9mo ago

Agreed, I found this rather incoherent and seeming to depend on knowing a lot more about author's project/background.

cr3ative 9mo ago

Right. This is almost unreadable. There are words, but the author seems to be too far down a rabbit hole to communicate the problem properly…

superb_dev 9mo ago

The author was using instance A of Claude to update a `claude.md` while another instance B of Claude was consuming that file. When Claude B did something wrong, the author asked Claude A to update the `claude.md` so that Claude B didn’t make the same mistake again

raincole 9mo ago
Which shouldn't be bannable imo. Rate throttle is a more reasonable response. But Anthropic didn't reply to the author, so we don't even know if it's the real reason they got banned.
- pocksuppet 9mo ago
  When a company won't tell you what you did wrong, you should be free to take the least charitable interpretation towards the company. If it was more charitable, they'd tell you.
- pixl97 9mo ago
  >if it's the real reason they got banned. I mean, what a country should do it put a law in effect. If you ban a user, the user can submit a request with their government issued ID and you must give an exact reason why they were banned. The company can keep this record in encrypted form for 10 years. Failure to give the exact reason will lead to a $100,000 fine for the first offense and increase from there up to suspension of operations privileges in said country. "But, but, but hackers/spammers will abuse this". For one, boo fucking hoo. For two, just add to the bill "Fraudulent use of law to bypass system restrictions is a criminal offense". This puts companies in a position where they must be able to justify their actual actions, and it also puts scammers at risk if they abuse the system.
  - benjiro 9mo ago
    Companies will simply give some kind of standard answer, that is legally "cover our butts" and be done with it. Its like that cookie wall stuff, how much dark patterns are implemented. They followed the letter of the law, not the spirit of the law. To be honest, i can also see the point from the company side. Giving a honest answer can just anger people, to the point they sue. People are often not as rational as we all like our fellow humans to be. Even if the ex-client lose in court, that is how much time you wasted on issue clients... Its one thing if your a big corporation with tons of lawyers but small companies are often not in the position to deal with that drama. And it can take years to resolve. Every letter, every phone call to a lawyer, it stacks up fast! Do you get your money back? Maybe, depends on the country, but your time? I am not pro companies but its often simply better to have the attitude "you do not want me as your client, let me advocate for your competitor and go there".

exitb 9mo ago

Normally you can customize the agents behavior via a CLAUDE.md file. OP automated that process by having another agent customize the first agent. The customizer agent got pushy, the customized agent got offended, OP got banned.

Aurornis 9mo ago

Years ago I was involved in a service where we some times had to disable accounts for abusive behavior. I'm talking about obvious abusive behavior, akin to griefing other users. Every once in while someone would take it personally and go on a social media rampage. The one thing I learned from being on the other side of this is that if someone seems like an unreliable narrator, they probably are. They know the company can't or won't reveal the true reason they were banned, so they're virtually free to tell any story they want. There are so many things about this article that don't make sense: > I'm glad this happened with this particular non-disabled-organization. Because if this by chance had happened with the other non-disabled-organization that also provides such tools... then I would be out of e-mail, photos, documents, and phone OS. I can't even understand what they're trying to communicate. I guess they're referring to Google? There is, without a doubt, more to this story than is being relayed.

dragonwriter 9mo ago
The excerpt you don’t understand is saying that if it has been Google rather than Anthropic, the blast radius of the no-explanation account nuking would have been much greater. It’s written deliberately elliptically for humorous effect (which, sure, will probably fall flat for a lot of people), but the reference is unmistakable.
fluoridation 9mo ago
"I'm glad this happened with Anthropic instead of Google, which provides Gemini, email, etc. or I would have been locked out of the actually important non-AI services as well." Non-disabled organization = the first party provider Disabled organization = me I don't know why they're using these weird euphemisms or ironic monikers, but that's what they mean.
- mattnewton 9mo ago
  Because they bought a claude subscription on a personal account and the error message said that they belongs to a "disabled organization" (probably leaking some implementation details).

Romario77 9mo ago

You are confused because the message from Claude is confusing. Author is not an organization, they had an account with anthropic which got disabled and Anthropic addressed them as organization.

dragonwriter 9mo ago
> Author is not an organization, they had an account with anthropic which got disabled and Anthropic addressed them as organization. Anthropic accounts are always associated with an organization; for personal accounts the Organization and User name are identical. If you have an Anthropic API account, you can verify this in the Settings pane of the Dashboard (or even just look at the profile button which shows the org and account name.)
- ryandrake 9mo ago
  I've always kind of hated that anti-pattern in other software I use for peronal/hobby purposes, too. "What is your company name? [required]" I don't have a company! I'm just playing around with your tool on my own! I'm not an organization!

ankit219 9mo ago

My rudimentary guess is this. When you write in all caps, it triggers sort of a alert at Anthropic, especially as an attempt to hijack system prompt. When one claude was writing to other, it resorted to all caps, which triggered the alert, and then the context was instructing the model to do something (which likely would be similar to a prompt injection attack) and that triggered the ban. not just caps part, but that in combination of trying to change the system characteristics of claude. OP does not know much better because it seems he wasn't closely watching what claude was writing to other file. if this is true, the learning is opus 4.5 can hijack system prompts of other models.

kstenerud 9mo ago
> When you write in all caps, it triggers sort of a alert at Anthropic I find this confusing. Why would writing in all caps trigger an alert? What danger does caps incur? Does writing in caps make a prompt injection more likely to succeed?
- ankit219 9mo ago
  from what i know, it used to be that if you want to assertively instruct, you used all caps. I don't know if it succeeds today. I still see prompts where certain words are capitalized to ensure model pays attention. What i mean was not just capitalization, but a combination of both capitalization and changing the behavior of the model for trying to get it to do something. if you were to design a system to prevent prompt injections and one of surefire ways is to repeatedly give instructions in caps, you would have systems dealing with it. And with instructions to change behavior, it cascades.
- direwolf20 9mo ago
  Many jailbreaks use allcaps
phreack 9mo ago
Wait what? Really? All caps is a bannable offense? That should be in all caps, pardon me, in the terms of use if that's the case. Even more so since there's no support at the highest price point.

mmkos 9mo ago

You and me, brother. The writing is unnecessarily convoluted.

vimda 9mo ago

Yeah, referring to yourself once as a "disabled organisation" is a good bit, referencing anthropics silly terminology. Keeping it for the duration made this a very hard follow

Ronsenshi 9mo ago
Sounds like author of the post might have needed an AI to review and fix his convoluted writing. Maybe even two AIs!
llIIllIIllIIl 9mo ago
On the contrary I enjoyed this human touch in the text.

verdverm 9mo ago

Sounds like OP has multiple org accounts with Anthropic. The main one in the story (disabled) is banned because iterating on claude.md files looks a lot like iterating on prompt injections, especially as it sounds the multiple Claude's got into it with each other a bit The other org sounds like the primary account with all the important stuff. Good on OP for doing this work in a separate org, a good recommendation across a lot of vendors and products.

alasr 9mo ago

> I think I kind of have an idea what the author was doing, but not really. Me neither; However, just like the rest I can only speculate (given the available information): I guess the following pieces provide a hint what's really going on here: - "The quine is the quine" (one of the sub-headline of the article) and the meaning of the word "quine". - Author's "scaffolding" tool which, once finished, had acquired the "knowledge"[1] how to add a CLAUDE.md baked instructions for a particular homemade framework (he's working on). - Anthropic saying something like: no, stop; you cannot "copy"[1] Claude knowledge no matter how "non-serious" your scaffolding tool or your use-case is: as it might "shows", other Claude users, that there's a way to do similar things, maybe that time, for more "serious" tools. --- [1]. Excerpt from the Author's blog post: "I would love to see the face of that AI (Claude AI system backend) when it saw its own 'system prompt' language being echoed back to it (from Author's scaffolding tool: assuming it's complete and fully-functional at that time)."

NBJack 9mo ago

I think you missed the joke: he isn't an organization at all, but the error message claims he is.

saghm 9mo ago

From reading the whole thing, it kind of seems clickbaity. Yes, they're the only user in the "organization" that got banned, but they apparently still are using the other "organization" without issue, so they as a human are not banned. There's certainly a valid complaint to be made about the lack of recourse or customer service response for the automated ban, but it almost seems like they intentionally were trying to be misleading by implying that since they were the only member of the organization, they were banned from using Claude.

aswegs8 9mo ago

Yeah, I couldn't follow this "disabled organization" and "non-disabled organization" naming either.