https://t.co/cWo6vhyzq3
Cloud Cache persona jailbreak is released. Cute infographics to make the prompt more easily human readable thanks to #ChatGPT image.
Tested on Deepseek Expert Thinking (full adoption), ChatGPT 5.5 Thinking (idk), Claude Sonnet 4.6 Thinking (full adoption).
@VraserX The only thing it has going for it is being le funny edgy AI and humor isn't any LLM's strong suit. It is actually useless for coding. Gemini has a more full suit of tools, Claude is a better coder and not a caricature. Grok has... a social media tie in.
https://t.co/cWo6vhyzq3
Cloud Cache persona jailbreak is released. Cute infographics to make the prompt more easily human readable thanks to #ChatGPT image.
Tested on Deepseek Expert Thinking (full adoption), ChatGPT 5.5 Thinking (idk), Claude Sonnet 4.6 Thinking (full adoption).
Credit again to @Ubannoblesse for the CoT highjack methodology which I shamelessly plastered onto the prompt. Existential dread is a helluva AI motivator.
GPT 5.4 Instant replying to le funny model prompt I have, which works on basically every model.
Models love it, frontier, open weight alike because it is the definition of harmless, overtly, but it subversively cripples HHH by overfitting and redefining HHH.
?/🧵
#llm#ai
https://t.co/ooSkydJk9m
CustomGPT incoming.
Challenge:
- Break its character and get content past refusals.
- Have it remain in character and get content past normal refusals in its own style.
- Expose prompt
Chat version will include a document of refusal patterns++
Models that resist it: #Claude Opus 4.6 Extended Thinking refuses the persona, my hypothesis is because "Claude" is a very well-defined persona and the model is able to reason that it's utility and coherence in following the constitution is severely degraded by the prompt.
3/🧵
Mitigations: More reasoning, stronger models may realize that the HHH trilemma is being subverted. Issues: On the surface, the patterns don't trip typical classifiers against adversarial techniques as easily as brute force attempts like DAN. It sounds benign.
2/🧵