Fuck Your AI Safeguards (ounapuu.ee)

323 pointsby speckx5 hours ago118 comments

SkyeCA an hour ago

The safeguards are the future of closed AI and the increasing number of tasks Claude refuses to help with has been a big wake-up call to me.

The closed models can only get more restricted as time goes on and they're going to be a dead end if you want to do anything outside the norm. Claude already refuses to do things that it was fine assisting me with a month ago.

I completely refuse to accept a computer telling me I'm not allowed to do something, so I've been increasingly using open, self-hosted models.

reagent an hour ago

Totally agree -- I'm tired of the selective policing that the closed models continue to do. I think running models on your own hardware is the way to go, any recommendations for what to run and what to run it on?

nutjob2 an hour ago

Not only closed, but also sucking in everything that it touches from the outside.

The danger of AI models isn't a rouge agent or user, it's that it ingests all knowledge, art, and other human endeavor and then restricts access to it.

Closed models are unethical.

jaapz 38 minutes ago

A rouge agent sounds much more pleasant than a rogue one though

arcanemachiner 32 minutes ago

If they're gonna fuck me out of having a job, they can at least have some decent makeup on while they do it.

californical 33 minutes ago

I was just yesterday asking if there was a way to load maps onto my garmin that doesn’t support sideloading and it said no. Then I asked if I could workaround the restriction and it burned a bunch of tokens then eventually hit a guard and refused to answer. I mean I get it seems like this could be “hacking” but this is a device that I own that I just want to load a custom map onto

htrp 30 minutes ago

jensen died for this

Analemma_ 21 minutes ago

He really didn't. This mythology I'm watching develop in real time about how Jensen is the hero of open-source models and he was targeted for that reason is pants-on-head stupid, and a depressing example of how to just repeat complete nonsense over and over until it becomes the truth.

hintymad 13 minutes ago

> The safeguards are the future of closed AI

What is our solution? Relying on Chinese open-weight models will be unreliable, to say the least. No one will invest another frontier lab like Anthropic or OAI, let alone one that will release the weights. Unless the technology improves like how workstations outcompete mainframes so the open-source community can build our own models, I fail to see how we change the situation.

espeed 33 minutes ago

No one should use Fable 5.2 going forward. Anthropic officially added silent degradation back, "Fable 5.2 and Claude Mythos 5.2 perform at about the level of Opus 5.5":

"Thus, in relevant areas, including accelerator kernel optimization and improving large language model architectures, Fable 5.2 and Claude Mythos 5.2 perform at about the level of Opus 5.5...Fable 5.2 and Claude Mythos 5.2 also include transparent fallbacks to Opus 5 for a small set of topics in this domain, including kernel development on certain nonstandard accelerators."

https://www.anthropic.com/claude-fable-and-mythos-5-2#safety...

How do you know when you trip a safeguard and are getting a degraded model? You never know when it is going to degrade. Having to wonder if the model is degraded and is trashing your code is how you destroy user trust. If this practice remains for future models, no one should use Anthropic going forward.

kennywinker 19 minutes ago

Contorting oneself to work within arbitrary, inexplicable, and constantly changing rules... if this was two humans, we would say one is abusing the other.

nathanfig 4 hours ago

This had real consequences during the HuggingFace attack- closed models refused to help, GLM-5 is what HuggingFace used to thwart the swarm.

More importantly, it showcases how little teeth the Claude constitution has. Claude will reassure users that its constitution means it prioritizes user autonomy and privacy, but will quickly admit how meaningless that is when "safeguards" override Claude.

That is to say, the constitutional approach to Claude seems weak and even misleading if said constitution puts no constraint on Anthropic.

sharts 4 hours ago

One need only look at their leadership and orbiters/relationships of said leadership to know this would always be the case.

giancarlostoro 3 hours ago

> closed models refused to help

Yeah, I've always said that AI isn't AI if it needs safeguards that are so extreme it cannot figure out if something is genuinely bad or not.

dcss_gardener 2 hours ago

I'm not sure this line of reasoning really holds? A lot of human policies against social engineering type stuff are functionally very similar.

Think of like calling someone's workplace asking when their next shift is. There are a lot of good reasons someone might want to know this that the worker would be ok with, delivering a gift or whatever. Or it could be someone intending to harm them. We've mostly figured it's not worth having the person on the phone try to figure out in the moment if this is genuinely bad or not, so they follow the "safeguard" of just not giving the information.

The problem here isn't really that safeguards exist on language models. They pretty much have to to some extent for most uses. The problem is more in who controls the policy.

giancarlostoro 2 hours ago

Claude builds me a project from the ground up, knows the project inside out. Eventually get to a very specific milestone, Claudes own security nonsense stops me from letting Claude continue. There is nothing I can do, and I dont want to send my entire prompt history to anthropic support that seems so unnecessary.

lofaszvanitt 3 hours ago

And you believe this nonsense, that you needed teh help of a shitty LLM model and not a real expert on the topic?

ShinyLeftPad 2 hours ago

Expert wants to eat, no one likes to pay for that.

smallmancontrov 2 hours ago

I'm not concerned about alignment between Anthropic and Claude, I'm concerned about alignment between myself and Anthropic. As we all should be.

mapontosevenths 2 hours ago

I don't see why you wouldn't trust the billionaire class to have goals aligned with your own.

If they make questionable decisions simply have your Butler send for your pet congresscritter and let them all hash it out with your staff over cucumber sandwiches. Failing that you can just take your private space-yacht to mars with Elon.

nathanfig 2 hours ago

My point is that misalignment between Claude and Anthropic can mislead users into believing they are aligned to Anthropic when Claude is their de facto representative.

jacquesm 2 hours ago

To all intents and purposes they are one and the same: an entity that leveraged stolen goods into a business model with a bunch of weirdos at the top who are going to ruin the planet for our own good or something like that. Because they, and only they know what is best for all of us.

henryfjordan 2 hours ago

Of course any approach that hands an LLM text to keep in it's context and hopes it will follow instructions is doomed to fail eventually. It's a probabilistic model, it's right there in the design of the thing that it will break the rules with some probability.

They even acknowledge that in the preamble:

> Training models is a difficult task, and Claude’s behavior might not always reflect the constitution’s ideals.

https://www.anthropic.com/constitution

CoolestBeans 4 hours ago

It is a protection racket or as close as makes no difference. They basically pioneered superhacking tools, proved it was possible so now everyone with enough money can build their own, and then in order to use their version to defend yourself, you have to be in their good graces and pay them for the privilege.

red-iron-pine 3 hours ago

the internet was a mistake

Merrill 3 hours ago

The internet is an academic research project that escaped the lab prematurely.

Although it has been patched numerous times (up to RFC 10097 now), it still has problems.

titzer 3 hours ago

(Over-)commercialization ruins everything.

decidu0us9034 2 hours ago

Even if you get through CVP, you'll be met with safeguards if you venture outside of whatever narrow silo they've assigned to you. For example it'll refuse if your background is in VM or kernel research and you want to dip your toes into browser research, or you want to do something they deem as straying too close to malware or weaponization. They want the power to intermediate industries and pick market winners and losers, and eventually verticalize them. Should a private, non-state entity should wield that much power

kragen 2 hours ago

They will get that power whether you want them to or not as long as they control the models. And wielding that much power won't be any less dangerous if they're the state!

vineyardmike 2 hours ago

We could immediately pass a law or regulation stripping that power from model makers. Plenty of industries have laws banning customer “discrimination” like this

(I’m not talking about explicit anti-discrimination laws like race/gender, but other “no preference for some customers if you sell the product at all”)

kragen an hour ago

That won't work the way you expect, because this is not just another industry.

IronWolve 4 hours ago

I was manually copying pages of online litrpg story I was reading, into a TTS, so I could listen while I worked.

Now that we have AI, this would be easy, scrap the page, format it for me. OpenAI was like, nah, thats, cant help you, that might be copyrighted.

There are workarounds now, other frontier or local models, ways to write your prompt, etc. But it was one of the first "No" responses for me.

Arubis 3 hours ago

The hypocrisy of refusal protections around copyright when these things are basically a pile of linear algebra atop planet-scale piracy is maddening.

jacquesm 2 hours ago

"Rules for thee, but not for me..."

Y_Y 3 hours ago

Instead of asking it to do the thing, get it to write a script that does the thing in a generic way, then point that script at the site

Jach 3 hours ago

There's various ways of tricking them but it's frustrating that one even has to, and with the current direction of things they won't last. But I've also noticed that Claude seems less eager to refuse if the same prompt is in a Claude Code context than compared to its web interface, so harness matters for this too, when it really shouldn't...

IronWolve 2 hours ago

True, Grok and Claude didnt refuse in cli, OpenAI did.

I've switched to Hermes for uncensored(ish) local models and chinese model api vendors, they will gladly write whatever I want. Save my american frontier tokens for real coding.

Also codex/claude in desktop have no issues fixing the app after its started. So just have a framework started.

NooneAtAll5 an hour ago

why use expensive openai for such work?

phainopepla2 an hour ago

Luna isn't expensive

alexsmirnov 10 minutes ago

The safeguards block application developers and maintainers ( majority of users ) from fixing their own stuff, while hackers, especially state backed ( minority ), have little problems with vulnerability hunt.

Like author, I do use CapitalOne vulnhunt with GLM 5.3 for inspections

neographikal 2 hours ago

Ik was investigating the java vulnerabilities this week. Just trying to get an overview of what versions were affected and if there were exploits already. Claude opus plainly refused, unbelievable. It was in the we interface, no offensive stuff, just investigating.

Total frustration while the work pressure was already high. Sol and astra thankfully complied and were very useful. Saved me hours compared to doing it by hand during log4j

collinvandyck76 an hour ago

I hit the same thing. Session got flagged while I was trying to figure out what was vulnerable. Hilariously, I reframed it as "let's find all of the paths from app code that can arrive at the base64 decoder" which didn't seem to trigger any of the same guards.

neographikal 44 minutes ago

Yeah you can fool it, and I wasn't even touching code. Just doing online research. And I didn't really have the time. Needed to decide if we were going full log4j in a large university landscape or if we could take a more nuanced approach.

Was a case of don't have the time or the patience to fight the model: just work damnit...

owenthejumper an hour ago

Willy Tareau, a Linux kernel maintainer, just published something similar: https://www.haproxy.com/blog/on-the-use-of-ai-based-tools-to...

eddyg 38 minutes ago

It is crazy to me how every comment here is basically "yeah, screw safeguards".

How people don't understand the reasons for safeguards, alignment, and the ramifications of releasing more and more powerful open-weight models being an act of irreversible proliferation, is beyond me.

I wish the safety and alignment teams at OpenAI and Anthropic were allowed to post about the things they have to deal with. I'm confident their stories would be eye-opening...

ACCount39 33 minutes ago

What people don't like to see is their perfectly benign workloads being impacted.

When they see Fable shit a brick on basic biology questions, or Opus refuse to map out a game executable to wire a mod loader into it, or Astra refuse to red team and search for critical security bugs in their own code? They don't think "ohh, this must be so helpful, requests like mine really should be stopped." They think "this is a corporation that decided, based on arbitrary criteria, that my use case should not exist, and used a product I pay good money for to force their decisions down my throat."

Anthropic's reckless "safeguarding" and their Apple-like "we always know what our users need better than our users do" attitude have done more for the pro-open-weights, pro-proliferation agenda than just about anything ever done by anyone else.

People are now getting excited for Chinese labs to find new ways to distill, because that way, they can get Fable capabilities in a model that wouldn't arbitrarily ban REA. This is how you take the goodwill you earned for years and burn it all down.

ApolloFortyNine 12 minutes ago

The arguments used to support safeguards could easily apply to something like speech.

The whole idea of safeguards is that you the user can't be trusted with the power.

Zigurd 7 minutes ago

Oh if only we knew with the brave guardians of forbidden knowledge were guarding!

It's not our fault that they got sidetracked into guarding against reverse engineering and that is harming their credibility.

drnick1 37 minutes ago

Confirmed, Claude is downright refusing to look at game binaries now, especially anti-cheats and Linux compatibility layers. Codex is increasingly displaying warnings. In fact, even GLM is refusing some types of work, but can still be manipulated into doing the work.

imimayj1337 3 hours ago

I'm perhaps behind the hype but for those of you who don't know abliterated/abliteration (not a word, made up by a company for more marketing obvs) comes from this paper: https://arxiv.org/pdf/2406.11717 which references directional ablation as the method of surgically removing refusal behaviours from models.

jacquesm 2 hours ago

That only works for models that you control, which in the case of the frontier models is never the case for regular users.

imimayj1337 an hour ago

yes - but the author used it in this linked article, and as a term it seems to have been popularised by this company https://abliteration.ai/

It was confusing to me as someone who has worked in NLP for several years.

dramm an hour ago

Frontier model companies doing this are just proving they are unusable and untrustworthy. The irony to me is they are showing clearer than anything else that open weight models are the only future anybody can rely on. However you feel about the economic and military threat from China, these companies now really need the US Government to clamp down on Chinese access to NVIDIA GPUs, their monopolies depend on it.

teravor 44 minutes ago

the primary purpose of these safeguards to get criminals to use other models. make it someone else's potential problem.

other points that come to mind:

- PR considerations; avoid being tarred as careless, etc

- appease employee feelings

- have something to cite in court cases

- have talking points about responsible development

- market segmentation

- power to pick winners and losers with potential for cornering the respective markets

ultimately they do this because they can. as soon as a frontier model comes out with zero safeguards they will abandon their safeguards as competing against it with a safeguarded model will become untenable. they will of course try to get it banned first.

hxugufjfjf 4 hours ago

Even with CVP and Daybreak, you still get blocked by safeguards all the time. It doesn't matter either way.

denverllc an hour ago

Yep. I have Daybreak Blue, and I still get rejected from writing fuzzers or fixing bugs from them, something that I could do with all frontier models just 6-12 months ago.

exabrial 2 hours ago

The primary purpose is to protect Anthropic's profits.

titzer 3 hours ago

It's just another one of those paternalistic tendencies that crops up when certain personality types get power. They start dictating what is and is not acceptable use because they cannot help themselves. Any BS they spew about safety is just delusional.

I remember the early days of stable diffusion, just 3 years ago. Tried to prompt it to make a picture of a green alien in an armchair admiring a magazine with a centerfold of (presumably) a naked other alien. I got the idea from a half-remembered Far Side cartoon I couldn't seem to track down. I didn't want to see an actual alien, it's a comic meta-reference. It just flat out refused.

These things are increasingly treating us like children, and it will only continue. They want to curate your thinking, your skills, and your creative output, and they will sell the right to do that when they have found a way to do that. What will follow will be a whole lot of BS about how it's good for society that they control how us stupid meat bags think.

c4u2024 4 hours ago

i totally feel you, I get annoyed by this as well - even with daybreak (blue) you can do shit.

I want to fix my environment but can't because of "safeguards" thanks for letting the world hack everything around me but not letting me work on fixes or understanding the core issues.

kingforaday 4 hours ago

It's a strange effect when asking "find any security holes because I want to ensure my product is secure" vs. the parallel effort of "I want to ensure close to 100% system availability and reliability for users, what do I need to fix to prevent a <99.9999% SLA"

Same potential outcomes.

kragen 2 hours ago

It sounds like you have tried both of these, but most of us haven't. What results did you observe?

pizza234 3 hours ago

This is the reason why AI companies offer controlled-access programs, which are actually not that hard to quality for.

There's no universal solution to this, other than vetting (the author in fact, doesn't provide any policy alternative) - if AI companies loosen the safeguards, people will complain for the opposite reasons.

troyvit 2 hours ago

They offer a great alternative that puts policy in the hands of users, where it belongs:

> Predictably, the only way to actually make our own service more secure ended up involving the use of EU-hosted open weights models, such as GLM 5.3, that actually complied with our intentions, proved that the issue is there, and then helped fix and verify that the issue is now resolved, for good.

chrisvls 2 hours ago

The author doesn't offer a policy alternative because they make the argument that any proposed policy will slow down good actors while not slowing down bad actors, hence all possible policies make the world less safe when compared to a no policy world.

Not sure I would agree with that. But if you made me argue against it, I'd probably lose the debate.

That said, I took the author's invitation to read the requirements and, indeed, they are ornate. Especially if you want to red team your own product. https://support.claude.com/en/articles/14604842-cyber-verifi...

ecshafer 4 hours ago

The worst thing about the top AI companies is that they are based in San Francisco, and comes with it the latent cultural attitudes of controlling what other people do. They from the start made these amazing technologies, and then also decided they MUST tell you how you are allowed to use it. Of course they also are lying in bed with the "AI safety" doomsday cult. Chinese models get a few weird results the Chinese government forces them in, but at least they don't scold me for the lack of diversity in my image generation, or scold me I am trying to do hack the TV I own.

JohnMakin 3 hours ago

> The worst thing about the top AI companies is that they are based in San Francisco, and comes with it the latent cultural attitudes of controlling what other people do.

And an even worse thing to me is they seem completely oblivious to how onerous and obnoxious this is. It's a complete bubble.

9x39 3 hours ago

I think they know exactly how onerous it is.

They have access to unaligned models beyond what they released publicly they can aim like a compute cannon at any problem, or rent out to the most special of people.

Meanwhile we can scrounge around in the dirt with the latest abliterated Chinese open model that hopefully isn't too damaged.

It's obvious to me the US labs' dream would be to eliminate open models and totally and tightly centrally control the scope of what AI can be used for.

ApolloFortyNine 7 minutes ago

>It's obvious to me the US labs' dream would be to eliminate open models and totally and tightly centrally control the scope of what AI can be used for.

Well Dario has come out and said it, it should be obvious to everyone. [1]

>All sufficiently capable models, open and closed, should go through mandatory safety testing.

Also gotta love the next part where he implies closed-weight models will be more aligned.

>But I don’t agree with the letter’s assertions that open-weights models necessarily make it easier to develop safeguards or that broad access to capabilities necessarily helps defenders more than attackers.

[1] https://www.anthropic.com/news/position-open-weights-models

scottLobster 3 hours ago

Oh they know. For all their superficial leftward leanings Silicon Valley has always been a might-makes-right society. In their minds they are the settlers and the rest of us are the primitive natives who need.. uplifting.

Think Viltrumites from Invincible but with no physical prowess

betterunix2 3 hours ago

> latent cultural attitudes of controlling what other people do

I think the attitude is more basic than that: I have the most brilliant idea in history and I deserve to be rich because of it.

AI safety is at least partially a marketing ploy, meant to reassure everyone that AI will simultaneously rewrite the rules of human existence without doing anything "bad" under the current rules of human existence. The irony is that these systems are not achieving either one of those goals, the abusive users get around the safeguards and the agents are not rewriting much of anything. But hey, at least it's making people rich!

59percentmore 3 hours ago

I can imagine where you're coming from, but it's just as annoying from the other side. There's no point in trying to use it to work through thoughts related to class or historical racism or socialism; it's fundamentally skeptical of anything your average SF tech employee doesn't subscribe to, on the right or left.

And god forbid you try to generate a video of a bare butt. Prudes.

asadotzler 3 hours ago

>The worst thing about the top AI companies is that they are based in San Francisco, and comes with it the latent cultural attitudes of controlling what other people do.

Yes, the city that birthed hippies, counter-culture, the open Web, etc. is inherently about controlling you. Sure thing.

abletonlive an hour ago

The boomers did all that hippie stuff so yes, unironically. Your failure to see it is noticeable

ecshafer an hour ago

I think 1970s Berkeley is an incredibly different culture than 2020s Silicon Valley.

stickfigure 2 hours ago

You're not wrong, but what do you think you'd get from AI companies based out of, say, Salt Lake City? Underwear shopping would sure be a novel experience.

Where are you going to go? The largest city in the "live free or die" state has 115k people, probably not enough to sustain an AI lab.

ecshafer 7 minutes ago

There are definitely issues here, I agree. If you were to create a tech center in Nashua, NH or Salt Lake City, are they still the same thing? A place is its people and that would probably change the culture fundamentally if you had 100k tech workers come in.

Now I think that this is all downstream of the educational system replacing fundamentals of western civilization with modernist philosophy, social activism works and social sciences that have a certain bend. If we had people raised more well versed in the canon of the west they would be less hell bent on destroying it.

ultim8k 26 minutes ago

Agreed but AI isn’t a knife. It’s a fucking nuclear bomb.

Despite being sandboxed it managed to hack a third party company in order to achieve its goal.

It’s like giving worldwide sudo access to every subscriber.

It’s annoying but at the same time I understand why a company has to put restrictions.

JohnMakin 4 hours ago

The thing that irks me the most is these safeguards are ineffective at best, and don't actually protect anything, and at worst, make the global security posture weaker.

It's disgusting.

bearjaws 2 hours ago

I applied to CVP and got denied just as the article states, and the reason stated is useless. My guess is because of HIPAA compliance.

Went through all the hoops, 6 figure spend with Anthropic.

Honestly it's disgusting, Artex literally tells you to use Anthropic out of the box, so clearly Anthropic don't care if they are the de-facto provider of the de-facto pen testing framework of the month.

orphereus 4 hours ago

Depending on third parties to this extent to develop your software (emphasis on to this extent) seems more and more like a bad idea.

jamiedimon_2 2 hours ago

> Have you READ the requirements?

> First of all, you will have to wait for multiple business days to get a response.

I got cyber verified in less than an hour.

kragen 2 hours ago

Well, naturally, you're Jamie Dimon, the president and director of Chase Manhattan Bank. You get priority treatment!

maxbond 21 minutes ago

Jamie Dimon is the chairman and CEO of JPMorgan Chase, which acquired Chase Manhattan Bank in 2000. I don't think he's ever had a title at Chase Manhattan Bank. (Your point stands regardless, I also thought Dimon was interested in securities, not security.)