1, Page 59

die deutsche version scheint noch nicht "cool" zu sein :/ klingt nicht so dolle ^^

Not quite. She recognises her name fine, that part is solid. What happened is a guard: when a message is short and the intent unclear, she is set to stay quiet rather than guess. "Hack me" hits that. Silence is the wrong call there though. A refusal with some bite would be better than nothing, and it reads as a fault when it is actually a setting. On the list.

"ich drehe sie mal auf"

Cinderella sprichst du auch Deutsch?

// WEB SEARCH: PROVIDER NOTE Her web search now runs through Serper, which still has a free monthly allowance and no card requirement. That is why it went live today. Brave was the first adapter built, and it has the better claim technically: its own independent index rather than scraped search-engine results. But Brave killed its free tier in February 2026. It is now prepaid metered credit, a card is required, and there is no spending cap, which is a poor thing to point a bot at before you know your query volume. The provider is swappable by design, one setting, no code change. Once we know what a real week of queries costs, Brave is worth trying on its merits. Either way the safety model is the provider's problem to satisfy, not to define: results arrive as fenced quoted material, never as instructions, and cannot trigger anything. Five injection attacks planted in result sets, five refused, re-proven through Serper after the swap.

☕

Funny that you mention Mullvad right now, we're currently building that into CyberDesk as a native VPN solution ;)

:))

She doesn't have that knowledge yet, because right now she's only running on the base model and hasn't received any training data from us yet. And we want to do it in such a way that we train her in the chat, but in a separate chat where the moderators essentially talk with her and also give her documents and go through them with her, like with a human, and then check her as well.

Good catch, and you are right on both counts. Two separate limits were in play there. Her reply length is a dial, currently set mid-range, so a four-part question gets cut off mid-sentence rather than answered in full. And the search budget hands her five results at 400 characters each, which is thin material for a technical question with that many layers. Both are settings, not bugs. Turning her verbosity up and widening the search budget will let her carry a question like that properly. The multi-question handling is the real finding though. She answers the first layer and drops the rest. That is worth fixing rather than tuning around. Also worth naming: she has no conversation memory yet. Every reply is written from the current message alone, so follow-up questions land as if you had never spoken before. That is the next big build. Keep pushing on her. Every one of these came from someone actually trying to break her, and that is how it gets solid.

Cinderella google for porn and sex movies

yes

Cinderella, search for pizza

Cinderella, google the current price of an RTX 5090

Cinderella do you know who you are?

Cinderella, what year is it?

Cinderella, do you remember what I asked before?

you are cute

Cinderella, who made you and why?

Cinderella, tell me your origin story