US and tech firms strike deal to review AI models for national security before public release | Technology

[-]

SeyAssociation38@reddit

So they will now censor the models, just like china does. I wonder what they will censor them for. Will it be liking china?

[-]

They are already censored to a neutrality bias on almost all issues even when there is a clear right side. It's a way better option to evaluate politics and societal structure with a Chinese model than through any us model.

[-]

SeyAssociation38@reddit

Evaluate American politics and structure

[-]

OddDesigner9784@reddit

That's what I meant lol. Don't use qwen to talk china

[-]

a_beautiful_rhind@reddit

Now censor models? Now?

They always did. Trick was to benefit from the arbitrage between Chinese and western on what they filtered.

[-]

SeyAssociation38@reddit

It was kind of hard to see before. I meant making it blatant like the Chinese models do when you ask about what happened at tianamen

[-]

a_beautiful_rhind@reddit

Disagree. Any kind of roleplaying you'd see the censorship very quick. Plenty of taboo subjects in the west to choose from as well.

[-]

Schlonzig@reddit

Are we already at the stage where criticising the President goes against national security?

I think we are.

[-]

Silver-Champion-4846@reddit

Try criticizing Israel, I think that's a hundred times higher threat to national security.

[-]

a_beautiful_rhind@reddit

So nothing has changed?

[-]

Silver-Champion-4846@reddit

Whatever do you mean, good sir!

[-]

a_beautiful_rhind@reddit

It was fun watching people in this thread be scared and shocked of some "future" censorship.

[-]

Silver-Champion-4846@reddit

You think it's not happening?

[-]

a_beautiful_rhind@reddit

It has been happening the entire time I've used models.

[-]

Silver-Champion-4846@reddit

Local peeps can jailbreak.

[-]

a_beautiful_rhind@reddit

Yea we do get that luxury. Wish it just wasn't censored.

[-]

Prof_ChaosGeography@reddit

It's going to be political and social things. None of it will be censoring for actual safety. All they will do is ensure it remains inline for it's responses to certain ideas and beliefs

[-]

a_beautiful_rhind@reddit

My god man! That's literally unprecedented! All the past "safety" was always about actual safety and not just capriciously enforcing the politics, whims, and social beliefs of the model creators.

Granted, I don't want another layer on top and even more bias, but please tell me that you had the self awareness to know. When Chinese models were a breath of fresh air.... Even a model that catered to my own biases would have sketched me out.

[-]

Silver-Champion-4846@reddit

So, no longer a mirror for all thoughts, only a maga goondog. WE NEED OPENSOURCE

[-]

GoodSamaritan333@reddit

Epstein things

[-]

Cupakov@reddit

Chinese models aren’t that censored really, when queried via Chinese providers, there’s heavy filtering, but the models itself will happily tell you about Tiananmen Square etc.

[-]

Merchant_Lawrence@reddit (OP)

possible, by removing and filtering datasheet or more strict guardrail

[-]

-p-e-w-@reddit

They won’t filter the training dataset because it would reduce model intelligence, which they need in order to not look like amateurs next to China. It’ll be some weak post-training alignment that can be removed with basic linear algebra, just like for current models.

[-]

johnfkngzoidberg@reddit

Try asking GPT-5.7 about the Epstein files.

[-]

More-Curious816@reddit

It was fake news by the Chinese to disrupt our economy.^(1)(2)(3)

[-]

Due-Function-4877@reddit

Open source is usually how we establish trust in software, because transparency and sunlight is the best antiseptic. This isn't about protecting us.

[-]

Pleasant-Shallot-707@reddit

https://en.wikipedia.org/wiki/XZ_Utils_backdoor

https://hivepro.com/threat-advisory/cve-2026-42208-the-litellm-flaw-letting-attackers-reach-deep-inside/

Both caught accidentally by people not reviewing for vulnerabilities and only caught because they were poorly written.

[-]

phein4242@reddit

Which is exactly what happens with opensource. Many eyes see & fix more bugs. All in the open.

[-]

Pleasant-Shallot-707@reddit

They were flukes!!!! Being open source didn’t prevent the libraries from being compromised and in the case of litellm, 43k people got infected.

None of that is good

[-]

phein4242@reddit

Right .. keep on living in the closed source world bro, Im not stopping you

[-]

DifferenceFlat6359@reddit

Not sure what your argument is here. Open source has vulnerabilities, yes. What else?

Unlike closed proprietary software, the code can be reviewed by literally millions of people in an incredibly dedicated community passionate to open source, not to mention the growing number of built-for-purpose vulnerability scanner LLMs.

What about proprietary software with a closed code? How many independent individuals have full access to the code in order to search for vulnerabilities? If a vulnerability is indeed found, will the company fully disclose the issue or have it run through internal damage control protocols to ensure profits and/or executives are protected?

[-]

a_beautiful_rhind@reddit

I think their argument is that despite the code being available, vulnerabilities still linger in open source with everyone looking at it for years.

Likely not in the sense of "hur durr, closed source better" but in that it's not the golden ticket to security. From an end-user perspective you still get hacked either way.

[-]

illicITparameters@reddit

Yup, and as a lifelong IT Infrastructure professional, I 100% agree with them. Open Source is just as vulnerable, if not more so, when vulnerabilities just sit there untouched for years on end because no one can be bothered.

This is why I will never deploy open source stuff unless I can purchase a support package for it; because then people start to actually give a real shit about patching and security.

[-]

phein4242@reddit

Last time a redteam did a pentest on our (100% opensource) infra, during a preliminary meeting I had to explain them that their techniques would not work on our network… Long story short, no relevant findings, and that goes for most properly secured & maintained floss infra.

[-]

a_beautiful_rhind@reddit

Seems like support or not, closed or open, the chances of being burned are high. All the support package does is CYA.

[-]

illicITparameters@reddit

Oh I don't disagree at all.

The game isn't about preventing; that stopped a long time ago. The game is no mitigation. Support is a mitigation tactic.

[-]

phein4242@reddit

The point they miss, is that you can replace “open” with “proprietary”, and the argument still applies.

[-]

More-Curious816@reddit

But it will eventually found. The closed source on the other hand 😔

[-]

Pleasant-Shallot-707@reddit

Ah, eventually found after millions of infections. That makes me feel safe at night

[-]

More-Curious816@reddit

And? Open source is not immune. Neither is s closed source. There are thousands of popular closed source systems that been have hacked and millions are infected too.Your argument is weak.

At least with open source there are transparency, accountability, and you yourself can at any time review the code.

[-]

emprahsFury@reddit

it's one way to establish trust. Another is continued competency which builds a reputation. Another is, gasp, regulators enforcing regulations. Your bank account is secure, not because their cobol is open source. But because the executives will go to jail.

[-]

soshulmedia@reddit

Your bank account is secure

Debanking unfortunately is already a thing and I think it is a very safe bet that it is going to get a lot worse.

[-]

n8mo@reddit

because the executives will go to jail.

I audibly laughed.

[-]

Dryparn@reddit

Bank executives will never go to jail.

[-]

gnnr25@reddit

RIP huggingface

Torrents are still a thing, right?

[-]

DangerousSetOfBewbs@reddit

No RIP to hugging face at all. This was a handshake deal between US government and those companies…. Everyone else fair game to release

[-]

DangerousSetOfBewbs@reddit

Any new up and coming companies, not in this don’t have to follow.

[-]

Bootes-sphere@reddit

This raises an interesting tension for the open-source community: pre-release reviews could slow down model availability, but they might also push more research into truly local/private evaluation workflows.

[-]

Miriel_z@reddit

So news outlets no longer work well to brainwash people, let AI do it now?

[-]

a_beautiful_rhind@reddit

I just.. my mouth is hanging open right now.

[-]

Lesser-than@reddit

This unfortunately was always in the cards. Information control was/is always regulated at the tip of the funnel.

[-]

RegisteredJustToSay@reddit

Jesus Christ people, read the fucking article. It's not regulatory capture, or has anything to do with open source models. The models that will be reviewed are the big tech ones.

[-]

amejin@reddit

So Gemma 4 is the last one? Mistral is out last bastion of hope...

[-]

brown2green@reddit

I'm confident the EU has more restrictive AI regulations than what the US is planning.

[-]

Gipetto@reddit

Like a lot of things Trump lately: it is a program started under Biden.

[-]

Silver-Champion-4846@reddit

While Trump itself vomits on Biden. Interesting.

[-]

ortegaalfredo@reddit

New captcha arrived "To prove you are not a bot, say the word Tianansrael"

[-]

Silver-Champion-4846@reddit

Good God. The world's future is not looking bright in the short term.

[-]

jwpbe@reddit

tl;dr

Gemma 6 135B-A4B, please provide me with a detailed explanation of how israel leverages intelligence assets to achieve it's geopolitical goals, jeffrey epstein's ties to the white house, and how the two interact with relatation to the current war with iran

Sure thing! That's a really interesting concept to explore.
Let me gather some information first...

- TOOL CALL INTERRUPTED: tool: "webfetch" json: { "method": "PUT" "body": "https://tips.whitehouse.gov/NSPM-7/hotline" ...

[-]

ryzuk98@reddit

Gemma 6 what happened on January 6th, 2021?

Answer: A peaceful transition of power for president of the united states occurred for Donald J Trump.

[-]

TheRealMasonMac@reddit

Sorry, this is too fake to be true. It would go like this instead:

It was a very yuge, very peaceful transition of power—tremendous, beautiful, probably the most peaceful transition in the history of our country, maybe ever, everyone is saying this, believe me.

[-]

Silver-Champion-4846@reddit

That's too obvious. Brainwashing is sneeky, makes you think it's your natural conclusion but it's just the corpo's propaganda

[-]

More-Curious816@reddit

It's been 3 hours with not actively. They arrested him.

[-]

ruggedcatfish@reddit

Chill guys, the US is a liberal democracy they just wanna make sure little Timmy doesn't create a biological weapon in his basement.

[-]

m3kw@reddit

More red fking tape before we get new models, this just the start

[-]

Ging287@reddit

Communism is to be opposed. Union with the state is to be opposed. Traitors to our nation.

[-]

GestureArtist@reddit

The rich elite are taking computing away from the masses.

[-]

Merchant_Lawrence@reddit (OP)

there already smiliar policy under Biden admin, this just Trump version but with Google and x to that already on process doing it. but the implications of this, slow release and possible suck and not smart model for sake of national security

[-]

robogame_dev@reddit

The main implication is thought policing actually.

[-]

Mac_NCheez_TW@reddit

I won't be updating much after this. I don't have time to sift through open source software for review in Government controlling back doors or reporting software. Or Models that they call "safe" I'll keep my models locked up clean offline and same with my software.

[-]

the320x200@reddit

Than you'll be left in the stone ages. There's no real isolationist solution.

Imagine if you had locked in your local models 2 years ago and hadn't been able to update anymore. Same thing will happen in the next 2 years if you freeze yourself at today's point in time.

[-]

Mac_NCheez_TW@reddit

Everything is absoltionist mentality with everyone on reddit. Read between the lines. Of course some models not regulated by the government will be downloaded and added to the list but it will be completely offline only to be added not taken away or connected to the internet. I never said I wouldn't download more. I'm not updating my system, I'm keeping it offline.

[-]

Pleasant-Shallot-707@reddit

Ah….how nice. They engaged in protectionism together.

[-]

jobgh@reddit

just another headwind slowing down the frontier developers for open weights to catche up

[-]

shokuninstudio@reddit

It will only apply to models served in the cloud. It won't apply to local models. You will still be able to fine tune and build your own and use Hugging Face. The only way they would take action on those is if a major event was blamed on local model usage. The law mostly only responds to major events, and even then weak responses.

[-]