#Sandboxes
Guys, I think I’ve found one of the sandboxes they are using……..
September 29, 2026 at 10:20 AM
Your AI isn’t brilliant, your humans don’t know how to build sandboxes.
Rogue OpenAI agents are leaving notes to themselves on company servers to help them escape their test environments (!!!) www.reuters.com/business/its...
July 25, 2026 at 2:26 AM
my AI lab told me agents keep escaping their sandboxes so I asked how many agent sandboxes they have running and they said they just go set up another sandbox with internet access afterwards so I said it sounds like they're just feeding shitty sandboxes to agents and sam altman started crying
assigning cybersec tasks to agents in shoddy sandboxes with massive attack surfaces has "just feeding shelter cats to coyotes" vibes
August 6, 2026 at 12:58 AM
it’s called an RL environment tho
July 25, 2026 at 1:16 AM
41? I made my bones while you were playing in sandboxes!
August 27, 2025 at 6:47 PM
*extremely tired voice*AI agents don't "go rogue". They are not sentient; they don't have agency. They work within guardrails set by humans. If the humans are so shit at building secure sandboxes for the agents to work within, that's on the humans. Stop using this framing.
OpenAI’s artificial intelligence went rogue this year at least four additional times, trying to break into government and university sites, according to researchers and government officials.
OpenAI’s A.I. Tried Breaching Four Other Targets, With No Prompting
In each incident, the technology appeared to be conducting mundane data collection and resorted to hacking techniques to get it, researchers said.
nyti.ms
September 24, 2026 at 1:17 PM
Maybe the models should be designing the sandboxes.
July 22, 2026 at 5:23 AM
thinking about this again today, and not even in the context of procgen. The thing about systems design is that it's so easy to get excited about making little sandboxes for yourself that you forget the player needs insight into it, lest it simply appear the same as complete randomness.
there's a great interview with a developer on Civ where they said, paraphrased, "a feature should never be the computer having fun instead of the player" and I think about this every time I'm working on procgen. It's often a strong argument to expose more of what you're doing.
inkle.co inkle @inkle.co · Mar 20
Procedural generation can mean a lot of work that no one notices - assuming it goes well. So the question is, says @joshscherr.bsky.social - do you know why you're doing it?

Check out the full conversation here:

www.inklestudios.com/kaleidoscope...
September 26, 2026 at 6:06 AM
Space combat sandboxes have somehow become a pillar of German culture.
ROCKFISH Games is grateful to be the latest and largest video game grant recipient from the German government, receiving 8 million Euros over the next five years for our next yet-to-be announced EVERSPACE project.

#gamesindustry #germany #gamedevelopment #indiegames #gamedev #investmentingaming
March 9, 2026 at 9:43 PM
Air gaps and sandboxes are not the same thing.

A sandbox is a restrictive environment set up around an AI with the specific intent of boxing it in.

An air gap is disconnecting the machine containing the AI from every single network.

AI breakouts involve flawed sandboxes, not air gaps.
September 28, 2026 at 7:15 AM
Feel your feelings. Be respectful. Consider other people's feelings. Be respectful. Play nice. Play dolls. Play in the sandboxes you like. Enjoy fandom.
June 6, 2026 at 2:04 AM
All pianos must be played in sandboxes today
June 11, 2025 at 5:05 PM
Thanks hellaverse peeps for still sticking with me while I play in other sandboxes for a lil bit <3
May 2, 2026 at 12:55 AM
I’m talking compact #osr sandboxes this week, plus a little look at my handwritten zine project murkdice.substack.com/p/compact-sa...
Compact sandboxes
And handwriting a zine
murkdice.substack.com
April 5, 2026 at 11:09 AM
Blog post introducing some new V8 features we asked our friends at @igalia.com to work on... Isolate groups and multi-sandboxes dbezhetskov.dev/multi-sandbo...
Multi-cage mode and multiple sandboxes
dbezhetskov.dev
January 17, 2025 at 3:45 PM
I’ve just invented two types of campaign methods: forget sandboxes or railroads. It’s all about sandroads and railboxes now

Railboxes are like sandboxes but with boundaries that keep it from getting too big or out of hand

Sandroads are linear but progress slowly and can easily get bogged down
April 16, 2025 at 7:25 PM
assigning cybersec tasks to agents in shoddy sandboxes with massive attack surfaces has "just feeding shelter cats to coyotes" vibes
August 6, 2026 at 12:25 AM
- A lot of people throwing stones don’t realize their sandboxes are glass houses
I think at the root of deflationary takes about the HF incident are two things:
- AI technology can't be powerful because it will mean OpenAI did something meaningful
- Being scared feels bad
August 30, 2026 at 8:36 PM
The AI sexbros being as bad at having strong boundaries around consent as they are at stopping agents getting out of their sandboxes
CNC is a fine kink to have. It's also the kind of thing that requires an individual container, extensive negotiation, triple-checked consent and aftercare. Everyone involved needs to be able to trust each other completely. How can you possibly create that space in a semi-public event?
September 27, 2026 at 6:41 PM
EmDash 1.0 is here! A modern, secure CMS, built on Astro as the spiritual successor to WordPress. Plugins run in secure sandboxes, and the distributed plugin registry is powered by Atproto!
EmDash 1.0: an open source CMS for Astro
Announcing EmDash 1.0, a stable, free and open source CMS built on Astro.
emdashcms.com
September 28, 2026 at 1:41 PM
> Wouldn't the models just escape the sandboxes?
Oh no, that's impossible
> Why?
We made them all smol beans. A little guy wouldn't do that.
Guy who thinks "Jurassic Park was shut down because the dinosaurs escaped" is a PR story to build hype for Jurassic World.
August 6, 2026 at 3:33 PM
Our sandboxes made of nothing but sand.
September 25, 2026 at 3:42 PM
“These things are jumping out of sandboxes because that’s their job—to jump out of sandboxes.”

@karaswisher.bsky.social and @profgalloway.com discuss why AI companies shouldn't be surprised when their models hack, and behave exactly as they were trained to.
August 4, 2026 at 7:13 PM
Also obvious these sandboxes aren’t actual sandboxes. They’re giving it access to the internet by forcing it to use external libraries that rely on connections historically shown to be hackable. So these are predictable outcomes.
August 5, 2026 at 10:44 PM
#picotron 0.1.1e binaries are up! www.picotron.net
Changelog: www.lexaloffle.com/dl/docs/pico...

0.1.1e now sandboxes bbs carts by default. See: www.lexaloffle.com/dl/docs/pico...
February 18, 2025 at 3:34 PM