It looks like whatever you try, the containment might be just temporary, agents successful escape their environments, using knowledge or finding new 0-days ¹

für mich irgendwie erschreckend, welche konstrukte erstellt werden um die "großen" weiter groß zu halten unter dem deckmantel von souveränität ¹

Pac-Man bench tests how well a model and harness can recreate Pac-Man from a single prompt: "Create a Pac-Man game in a single html page"

https://jonclegg.github.io/pacman-bakeoff/

what surprised me that deepseek is quite behind as well as the huge difference between the gpt model families (sol,terra,luna) - and also that kimi is so far behind.

e.g. https://github.com/omnichar/OmniChar

+

As I'm also part of the apple ecosystem since a while I'm as well quiet interested in what's going on. One of the major USPs I see, is the at least advertised end to end encryption for your cloud data, like photos documents and whatever you sync. What I didn't know: UK forced apple to change end2end encryption towards an encryption where they can offer access to government. Sounds like you should setup your Apple account outside of UK to be able to be protected or find other workarounds. I think governments who intrude their citizens privacy like that are failing.

https://gambit.security/blog-posts/autonomous-ai-agents-online-retailers-25-a-company

It's getting too cheap to run attacks on software, a recent report of automated agentic hacking discovers at least 600k stolen credit cards, from the report I'd guess it was a Chinese speaking people, but who knows. Mitigation seems hard considering the scale and experience modern ai harness can bring into such campaign. SQL filtering, web application firewalling, monitoring what's going on on the system, filtering filesystem status (e.g. are there scripts which should not be there?) might be helpful. 🙈

https://gambit.security/blog-posts/autonomous-ai-agents-online-retailers-25-a-company

First reports coming in that ai is fed with wrong information by attackers trying to trick users into going to wrong websites or calling fake numbers. Quiet interesting in context of these Google search, ai answer:

https://medium.com/@arielsimon/dark-sourcery-how-hackers-manipulate-ai-to-scam-you-88df434d2073

quite interesting: "Can gzip be a language model?" - https://nathan.rs/posts/gzip-lm/

took ~1hour to create this, and me casually fine tuning while watching movies or playing world of warcraft in the background for some hours on the weekend. this is the output of astra. all create by AI, even the deployment and api deployment on my vps. I only provided the ssh key and github PAT.

+

https://potato.minizap.online/

636606729769440499166579950236036751749912014371509557713570027508971809534551913252252094954941974952859310861988904737359709200557919 is a factor of RSA-896 Steve Weis (@sweis)

https://x.com/sweis/status/2101484464807596264?ref_src=twsrc%5Etfw

it feels like every week now a high end model with even more capabilities releases. openai astra is just some week old or so and they are about to drop new models while anthropic just sends opus 5.5 [^1] live into the leaderboards, of course on top of everything else - but openai just waiting around the corner with luna...

+ +

^[1]: https://www.anthropic.com/claude-opus-5-5

• zu spät: ✅
• spontaner Gleiswechsel: ✅
• klo defekt: ✅
• kinder dabei: ✅
• Anzeigentafel zeigt falsche Werte: ✅
• Anschlusszug verpasst: ✅

Ein typischer Freitag und die Reise ging gerade erst los.

You can adopt a pixel from the new space telescope on NASA by just providing your mail address here. https://science.nasa.gov/mission/roman-space-telescope/adopt-a-pixel/