zlacker

[parent] [thread] 21 comments
1. alista+(OP)[view] [source] 2026-01-22 19:20:58
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...

replies(8): >>lazyfa+C2 >>rvba+23 >>redeem+14 >>fallou+n6 >>Romari+K6 >>layer8+Fp >>healsd+kx >>crypto+Ys1
2. lazyfa+C2[view] [source] 2026-01-22 19:32:03
>>alista+(OP)
Author really comes off unhinged throughout the article to be frank.
replies(3): >>superb+a3 >>pjbeam+R3 >>static+46
3. rvba+23[view] [source] 2026-01-22 19:34:22
>>alista+(OP)
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.

replies(1): >>darkwa+ol
◧◩
4. superb+a3[view] [source] [discussion] 2026-01-22 19:34:50
>>lazyfa+C2
Did we read the same article? The author comes of as pretty frustrated but not unhinged
replies(1): >>ryandr+N8
◧◩
5. pjbeam+R3[view] [source] [discussion] 2026-01-22 19:38:04
>>lazyfa+C2
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.
6. redeem+14[view] [source] 2026-01-22 19:39:02
>>alista+(OP)
i have no idea what he was actually doing either, and what exactly is it one isnt allowed to use claude to do?
◧◩
7. static+46[view] [source] [discussion] 2026-01-22 19:48:13
>>lazyfa+C2
Author thinks he's cute to do things like mention Google without typing Google but I wouldn't call him unhinged.
8. fallou+n6[view] [source] 2026-01-22 19:49:28
>>alista+(OP)
This tracks with Anthropic, they are actively hostile to security researchers.
9. Romari+K6[view] [source] 2026-01-22 19:51:33
>>alista+(OP)
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.

◧◩◪
10. ryandr+N8[view] [source] [discussion] 2026-01-22 20:01:23
>>superb+a3
I wouldn't say "unhinged" either, but maybe just struggling to organize and express thoughts clearly in writing. "Organizations of late capitalism, unite"?
replies(2): >>Bootvi+xk >>genewi+AO
◧◩◪◨
11. Bootvi+xk[view] [source] [discussion] 2026-01-22 21:10:34
>>ryandr+N8
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.

replies(1): >>ryandr+Tn
◧◩
12. darkwa+ol[view] [source] [discussion] 2026-01-22 21:17:27
>>rvba+23
> 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).

◧◩◪◨⬒
13. ryandr+Tn[view] [source] [discussion] 2026-01-22 21:31:20
>>Bootvi+xk
I think if one's readers need an "ironic euphemism decoder glossary" just to understand the message, it could use a little re-writing.
replies(1): >>layer8+Wq
14. layer8+Fp[view] [source] 2026-01-22 21:40:26
>>alista+(OP)
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.
◧◩◪◨⬒⬓
15. layer8+Wq[view] [source] [discussion] 2026-01-22 21:46:59
>>ryandr+Tn
It was perfectly understandable to me. Maybe cultural differences? You seem to be American, OP Portuguese, and myself European as well.
replies(2): >>ashirv+wv >>superb+NT
◧◩◪◨⬒⬓⬔
16. ashirv+wv[view] [source] [discussion] 2026-01-22 22:18:31
>>layer8+Wq
Another European chiming in, I enjoyed OPs article.
17. healsd+kx[view] [source] 2026-01-22 22:29:11
>>alista+(OP)
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.
replies(2): >>Booris+IW >>riboso+R51
◧◩◪◨
18. genewi+AO[view] [source] [discussion] 2026-01-23 00:51:23
>>ryandr+N8
https://en.wikipedia.org/wiki/Late_capitalism

https://community.bitwarden.com/t/re-enabling-a-disabled-org...

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.

◧◩◪◨⬒⬓⬔
19. superb+NT[view] [source] [discussion] 2026-01-23 01:38:29
>>layer8+Wq
I’m American and it made sense
◧◩
20. Booris+IW[view] [source] [discussion] 2026-01-23 02:12:30
>>healsd+kx
The whole thing reads like LLM psychosis.
◧◩
21. riboso+R51[view] [source] [discussion] 2026-01-23 03:37:42
>>healsd+kx
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/9a0802af16f5a1ff1...

22. crypto+Ys1[view] [source] 2026-01-23 07:26:20
>>alista+(OP)
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.
[go to top]