Cisco’s study “Death by a Thousand Prompts: Open Model Vulnerability Analysis” shows that so‑called multi‑turn jailbreaks against open AI models (Open-Weight-LLMs) succeed in up to 92% of cases, revealing serious security vulnerabilities in their architecture. The text was automatically translated from German into English. The German quotations were also translated in sense.
I added a small notice that appears on posts older than a year. It was important to me that this notice not show up on photo-only posts — after all, photos rarely lose their relevance. First I wrote a working version myself, then had a language model further optimize the code. Now I’m running the following variant: {{- /* Tags prüfen */ -}} {{- $tags := default (slice) .Params.tags -}} {{- $tagsLower := apply $tags "lower" "." -}} {{- $hasPhoto := in $tagsLower "photo" -}} {{- /* Relevantes Datum bestimmen und Stichtag berechnen */ -}} {{- $pageDate := time.AsTime (or .Lastmod .PublishDate .Date) -}} {{- $oneYearAgo := (now).AddDate -1 0 0 -}} {{- /*Hinweis nur auf echten Seiten, mit gültigem Datum, älter als 1 Jahr und nicht beim Tag "photo" */ -}} {{- if and .IsPage (not $hasPhoto) (gt ($pageDate.Unix) 0) (lt $pageDate $oneYearAgo) -}} <div class="alert alert-warning"> Hinweis: Dieser Artikel ist über ein Jahr alt. Die Informationen könnten veraltet sein. </div> {{- end -}} The text was automatically translated from German into English. The German quotations were also translated in sense. ...
I learned something new! Although it’s almost obvious, it wasn’t clear to me. Right now we can’t simply fix faulty large language models when they make mistakes or spread false information. That’s because LLMs are created using neural networks. Neural networks learn on their own and build mathematical probabilities by repeatedly computing certain things with small variations. So if you want to change something in an LLM, you’d have to adjust countless such probabilities. This process is so labor-intensive that it’s easier to create a new LLM. That’s why GPTs are released as larger, major version updates. ...
In the future, UnRaid’s long-standing promise, “Buy once, use forever. No subscription. No hidden fees.”, might change. Unraid writes in »Upcoming Changes to Unraid OS Pricing« on unraid.net These will function similarly to the current Basic, Plus, and Pro licenses, with the main difference being that the Starter and Unleashed licenses will come with one year of software updates. After that, customers will be able to pay an OPTIONAL extension fee, which makes them eligible for another year of updates. If you choose not to renew, no problem. ...