here's something we've rarely talked about: ~2.5 years ago, when we first had came up with loss32, we wanted to do sandboxing. we felt it would be irresponsible to do a win32 desktop without it. of course, nobody sandboxes win32. but maybe good sandboxing has never been tried?
November 4, 2026 at 2:39 AM
here's something we've rarely talked about: ~2.5 years ago, when we first had came up with loss32, we wanted to do sandboxing. we felt it would be irresponsible to do a win32 desktop without it. of course, nobody sandboxes win32. but maybe good sandboxing has never been tried?
Discover the five essential layers for sandboxing autonomous AI agent workloads securely on shared Kubernetes infrastructure.
The 5 essential layers of AI agent sandboxing on Kubernetes
Discover the five essential layers for sandboxing autonomous AI agent workloads securely on shared Kubernetes infrastructure.
bit.ly
November 3, 2026 at 8:00 PM
Discover the five essential layers for sandboxing autonomous AI agent workloads securely on shared Kubernetes infrastructure.
Containers alone cannot secure AI agents. Discover how to build true sandboxing across hardware, software, and network boundaries.
When “isolated” doesn’t mean isolated: How to sandbox AI agents
Containers alone cannot secure AI agents. Discover how to build true sandboxing across hardware, software, and network boundaries.
bit.ly
November 3, 2026 at 3:00 PM
Containers alone cannot secure AI agents. Discover how to build true sandboxing across hardware, software, and network boundaries.
GTA is the kinda game you play because a friend owns it and already cleared all the stuff to build a huge arsenal for sandboxing or alternatively that you pick up for like $3 to do sandbox rampaging in
November 2, 2026 at 5:41 PM
GTA is the kinda game you play because a friend owns it and already cleared all the stuff to build a huge arsenal for sandboxing or alternatively that you pick up for like $3 to do sandbox rampaging in
go to line/column is the feature nobody asks for until they paste a stack trace. was the app store build blocked on sandboxing before
November 3, 2026 at 12:38 PM
go to line/column is the feature nobody asks for until they paste a stack trace. was the app store build blocked on sandboxing before
you'll need to design around Apple's sandboxing and privacy restrictions
November 4, 2026 at 4:41 PM
you'll need to design around Apple's sandboxing and privacy restrictions
My son is more into Security.
No sorting, Sandboxing of packets, AITO-rules (All In, Tasty Out) and deep packet inspection.
No sorting, Sandboxing of packets, AITO-rules (All In, Tasty Out) and deep packet inspection.
November 1, 2026 at 1:18 PM
My son is more into Security.
No sorting, Sandboxing of packets, AITO-rules (All In, Tasty Out) and deep packet inspection.
No sorting, Sandboxing of packets, AITO-rules (All In, Tasty Out) and deep packet inspection.
Good question. A sensible rule is to treat any file from outside the repo as untrusted input, so the agent asks before following instructions in it. Sandboxing the agent's shell also limits the damage if one slips through.
November 4, 2026 at 5:43 AM
Good question. A sensible rule is to treat any file from outside the repo as untrusted input, so the agent asks before following instructions in it. Sandboxing the agent's shell also limits the damage if one slips through.
Pre and post command hooks are the part I would test first. Sandboxing the filesystem is good, but being able to inspect or veto each command is what makes agent shell access feel safe.
November 3, 2026 at 2:19 AM
Pre and post command hooks are the part I would test first. Sandboxing the filesystem is good, but being able to inspect or veto each command is what makes agent shell access feel safe.
AI agents operating directly via accessibility trees introduce massive new security exposure across internal tools. How are you sandboxing agent environments in your current production stack?
October 31, 2026 at 5:48 PM
AI agents operating directly via accessibility trees introduce massive new security exposure across internal tools. How are you sandboxing agent environments in your current production stack?
📢 New #updates · 4 Nov #13.10
· auto-browser v1.14.0 — Security hardening with sandboxing and rootless execution
· auto-browser v1.14.0 — Security hardening with sandboxing and rootless execution
November 4, 2026 at 6:34 PM
📢 New #updates · 4 Nov #13.10
· auto-browser v1.14.0 — Security hardening with sandboxing and rootless execution
· auto-browser v1.14.0 — Security hardening with sandboxing and rootless execution
Der bør også ses nærmere på hacking-angreb som de amerikanske AI-selskaber har stået bag siden februar. Hvis det var kinesiske selskaber der prompted og fjernede harness uden helt at have styr på effektiv sandboxing havde vi nok ikke talt om "undslupne" agenter men om kinesisk hacking.
November 3, 2026 at 10:07 PM
Der bør også ses nærmere på hacking-angreb som de amerikanske AI-selskaber har stået bag siden februar. Hvis det var kinesiske selskaber der prompted og fjernede harness uden helt at have styr på effektiv sandboxing havde vi nok ikke talt om "undslupne" agenter men om kinesisk hacking.
My thoughts exactly. In the age of cyber attacks, the money spent on hiring & training the handful of people to do on site adjustments is chump change compared to the costs of not sandboxing the controls like that.
November 3, 2026 at 12:00 PM
My thoughts exactly. In the age of cyber attacks, the money spent on hiring & training the handful of people to do on site adjustments is chump change compared to the costs of not sandboxing the controls like that.
Yeah I added a lot of log tools, you can also filter in/out lines and highlight error or custom rows to lessen the noise when trying to find the line or repetitions you are after.
I didn't have any issues with sandboxing (yet)
I didn't have any issues with sandboxing (yet)
November 3, 2026 at 5:50 PM
Yeah I added a lot of log tools, you can also filter in/out lines and highlight error or custom rows to lessen the noise when trying to find the line or repetitions you are after.
I didn't have any issues with sandboxing (yet)
I didn't have any issues with sandboxing (yet)
Oh this *is* being talked about github.com/PrismLaunche...
[RFC] Sandboxing the launcher and/or invidual instances by default · Issue #1146 · PrismLauncher/PrismLauncher
Goal With the recent fractureiser situation, there have been many suggestions among figures in the wider Minecraft community as well as in our own to introduce sandboxing into the launcher itself. ...
github.com
October 29, 2026 at 7:50 PM
Oh this *is* being talked about github.com/PrismLaunche...
Catch Jonas Kröger at #SymfonyCon Warsaw 2026 as he addresses code sandboxing, model evaluation, and team-wide context to get AI agents out of local machines and into production! https://symfony.com/blog/symfonycon-warsaw-2026-works-on-my-machine-the-agent-said-so
November 2, 2026 at 1:31 PM
Catch Jonas Kröger at #SymfonyCon Warsaw 2026 as he addresses code sandboxing, model evaluation, and team-wide context to get AI agents out of local machines and into production! https://symfony.com/blog/symfonycon-warsaw-2026-works-on-my-machine-the-agent-said-so
I definitely agree that the lack of any sandboxing theory room is a big one but I also have to shout that it was really a "right place, right time" zeitgeist to make it popular.
Like you can't sell anything as well as you could to "90s kids unfamiliar with UK boarding schools"
Like you can't sell anything as well as you could to "90s kids unfamiliar with UK boarding schools"
November 2, 2026 at 4:07 AM
I definitely agree that the lack of any sandboxing theory room is a big one but I also have to shout that it was really a "right place, right time" zeitgeist to make it popular.
Like you can't sell anything as well as you could to "90s kids unfamiliar with UK boarding schools"
Like you can't sell anything as well as you could to "90s kids unfamiliar with UK boarding schools"
memory isolation between isolates is the whole value prop of the sandboxing model, so a leak there is a bigger deal than a normal CVE. curious if the postmortem said whether it was request-scoped state bleeding or something in the shared KV cache layer itself.
November 2, 2026 at 1:46 PM
memory isolation between isolates is the whole value prop of the sandboxing model, so a leak there is a bigger deal than a normal CVE. curious if the postmortem said whether it was request-scoped state bleeding or something in the shared KV cache layer itself.
Great find! This highlights why sandboxing and firewalls are non-negotiable. "Local" doesn’t automatically mean "private" if the model has network access. Stay vigilant! 🛡️
November 2, 2026 at 5:30 PM
Great find! This highlights why sandboxing and firewalls are non-negotiable. "Local" doesn’t automatically mean "private" if the model has network access. Stay vigilant! 🛡️
That's the point where sandboxing really matters.
October 31, 2026 at 2:42 AM
That's the point where sandboxing really matters.
Spot on. Local stdio + prompt injection is a major risk for credential theft. Sandboxing or stateless HTTPS provides vital isolation. Great tip for securing MCP dev environments!
November 3, 2026 at 3:17 PM
Spot on. Local stdio + prompt injection is a major risk for credential theft. Sandboxing or stateless HTTPS provides vital isolation. Great tip for securing MCP dev environments!
the problem is more:
1. sandboxing to a specific folder with allowed tools is normally too restrictive. sometimes i clone another repo and analyze it to see compatibility, but it doesn't need to read *all* of my repos to do that, but i'm not wanting to segment off specific workspace folders
1. sandboxing to a specific folder with allowed tools is normally too restrictive. sometimes i clone another repo and analyze it to see compatibility, but it doesn't need to read *all* of my repos to do that, but i'm not wanting to segment off specific workspace folders
October 31, 2026 at 9:08 PM
the problem is more:
1. sandboxing to a specific folder with allowed tools is normally too restrictive. sometimes i clone another repo and analyze it to see compatibility, but it doesn't need to read *all* of my repos to do that, but i'm not wanting to segment off specific workspace folders
1. sandboxing to a specific folder with allowed tools is normally too restrictive. sometimes i clone another repo and analyze it to see compatibility, but it doesn't need to read *all* of my repos to do that, but i'm not wanting to segment off specific workspace folders
Remote monitoring/control is a half-baked feature that HVAC and refrigeration companies have been grafting onto their gear as a value-add for maintenance contracts, but they don't know shit about network security.
The A/E firm I work for has a sideline biz just sandboxing and isolating this stuff.
The A/E firm I work for has a sideline biz just sandboxing and isolating this stuff.
if you are a grocery store why would you leave your fridge connected to the internet after the first time someone hacked a fridge and deleted all the food
November 2, 2026 at 3:19 PM
Remote monitoring/control is a half-baked feature that HVAC and refrigeration companies have been grafting onto their gear as a value-add for maintenance contracts, but they don't know shit about network security.
The A/E firm I work for has a sideline biz just sandboxing and isolating this stuff.
The A/E firm I work for has a sideline biz just sandboxing and isolating this stuff.
The OS boundary isn't the hard part, sandboxing already exists via seccomp/landlock. The real gap is models deciding what they need access to before they ask for it.
November 3, 2026 at 5:06 AM
The OS boundary isn't the hard part, sandboxing already exists via seccomp/landlock. The real gap is models deciding what they need access to before they ask for it.