btw what does it say about the AI safety nonprofit circuit that after years of funding, conferences and papers, nvidia has to turn up with an actual reference design? you lot have described the agent escaping in extraordinary detail. was anyone assigned to build the fucking enclosure, or were you waiting for a philosopher to define “outside”? jensen is trying to sell the machines and somehow he’s ended up doing your homework too.
Jensen Huang
@JensenHuang
Sep 28
Replying to @JensenHuang
NVIDIA Open Agent Safety Platform Reference Design combines NVIDIA OpenShell and NVIDIA Sentry. OpenShell is an open-source secure runtime that gives AI agents clear, enforceable boundaries. It traces their actions and enforces policy as they work. NVIDIA Sentry delivers added layer of security with hardware-based enforcement on NVIDIA BlueField, continuously monitoring agent activity through a trusted telemetry and detection pipeline and enabling millisecond-scale containment and quarantine.

Sep 28, 2026 · 6:01 PM UTC

22
48
491
19,854
Sort replies: Relevant Recent Liked
Replying to @basedjensen
This is a sandbox architecture. It’s not a solution to ai safety. Its a start and the easiest bit.
1
1
361
Yea but they haven't done this either after years and billions spent
1
25
309
more replies
Replying to @basedjensen
Alignment and sandboxing aren't the same thing or substitutes.
3
6
390
You need sand boxes for testing the whole reason we have these "hacking incidents" is because they were runing alignment research on insecure sand boxes
1
31
344
Replying to @basedjensen
Solving the problem puts them out of a job and cures rheir neurosis! Can't have that.
16
419
Replying to @basedjensen
This is embarrassing for them. Please we that research how students can learn better math & sciences with AI (as tool, that’s what they are) shouldn’t be confused with them. We do have empirical measurements and testable hypothesis. Thank you for your attention in this matter!
1
331
Replying to @basedjensen
Not to mention they’ve trapped themselves. How will it look to run on hardware that doesn’t do stuff like this now? Jensen is taking full advantage of the business inexperience
7
297
Replying to @basedjensen
The parents doing the kids homework lmao 😂
1
144
Replying to @basedjensen
Incentives…
1
273
Replying to @basedjensen
Their solution is a government agency that centralizes control and power that they get to lead.
33
Replying to @basedjensen
they just can't
generatorman@generatorman_ai
Aug 9
something curious about ai risk people is that they never ever talk about defence. they can imagine swords, and they can imagine sheaths (which they do obsessively), but they are unwilling to imagine shields. i wonder why.
3
124
Replying to @basedjensen
Exactly. A strong sandboxing buys you enough time to do alignment research. Unfortunately, they would rather do doomerism instead of hard work.
25
Replying to @basedjensen
This is so true 😅 It's called platform engineering, and this is admittedly at massive scale, but yeah, daddy has to step in apparently 😅🤦
1
142
Replying to @basedjensen
Jensen right now:
3
93
Replying to @basedjensen
lol
1
179
Replying to @basedjensen
took me 10 mins with AI to design a "safety container" still have to iterate a bit and stress test it but still.
1
1
234
Replying to @basedjensen
Have you ever seen an NGO actually try to solve the problem they ask money for?
6
90
Replying to @basedjensen
@pangram ai?
1
2
79
Replying to @basedjensen
The purpose of a system is what it does. The AI safety people are not interested in practical solutions that will contain the current models because this run contrary to their long term beliefs that superintelligence can not be contained. Plus if the main source of fear disappears, how are they going to continue getting funding. There is no world in which solving the problem is a good outcome for them. Pause and slow down on the other hand means the threat is ever present and the need of vigilant watchers is permanent.
33
Replying to @basedjensen
68
Replying to @basedjensen
lmfao
2
159
Replying to @basedjensen
I can’t believe this but in the past few days I’ve decided I like Nvidia. At least they’re trying to help people build tech.
3
98
Replying to @basedjensen @curtis_yarvin
Just don’t build a superintelligence
36