Please regulate us, now! If we don’t get regulatory capture ASAP I’m toast. I mean look man I just had to go beg for tens of billions more cash so I can keep the lights on and my IPO is looking more and more screwed every day. Please, I’m begging you. What more do I need to say to convince you?
Whenever there is an adverse event, people say of course it should be regulated. The EU AI Act is a good start but we need more. These labs need to be held accountable just like all industries that can impact others. Regulate the chips, regulate the models, regulate the companies.
When these companies ask for regulations, people say they are trying to do regulatory capture. They are trying to ban opensource models, and cement their place as the AI arbiters and ban all others. We need competition and open weight models and China is leading the charge.
I understand that the 2 position aren't exactly opposed. We probably need regulations, and ideally ones that don't produce regulatory capture. That is extremely hard though. Almost all regulations end up creating a form of regulatory capture. Any regulation compliance has a cost. The more complex and wide reaching the regulations are, the more expensive it's to comply and the more they cement the existing players and make it harder for new ones to enter the market. FAA regulations makes it so we don't have dozens of passenger airplane startups that are trying crazy wacky things. Food services are regulated too, but the compliance is fairly doable by mom and pop stores, yet probably too high for your unlicensed pop-up taco stand.
I haven't really seen any meaningful discussion about what AI regulation should be. Only extremes of "No regulations, all is fair" or "no one should have access to running or developing models but OpenAI and Anthropic".
So yeah, US regulation is ONLY about capturing US consumers. This isn't food or airplanes.
What Altman, Dario and other such folks are trying to do is put a price on getting the model audited, by playing on these: price, timelines, red-tape.
Maybe a test-suite, kept continuously updated, similar to those benchmarks we already have for measuring the intelligence of these models. Like arc-agi, or terminal bench, or more.
Sure, research will continue in secret, just like nuclear, biological and chemical weapons. It's just that there's going to be an economy backing it up, with MAD ensuring the attacks are "civil", as Axl once said.
I’m explicitly criticizing that position.
Seems like you're trying to move the discussion towards the tradeoffs of more vs less regulation, both of which are "moderate" talking points.
Your "extremes" are "free for all" and "nobody but big tech"
Here's a true extreme: NOBODY
But the reality is that right now it's an arms race and the only countries crying for regulations are the ones that know they've already lost.
So I would agree that this is bringing more of this to light, but the idea that unknown exploitable bugs can get weaponized being a new thing is naive.
If anything putting all this out in the open makes things more secure, not less, in the long run as it’s harder for hackers to know and exploit them before they’re known as they don’t survive undetected for long. So a bit painful in the short term but we get more secure infra in the long term.
Still not a reason to give Sam regulatory capture as an out to saving OpenAI.
Nobody of course, this is America!
This is the textbook definition of racketeering.
In any case, what I want to say is this: All technology has tradeoffs but this has shown we are locked in a game theory mechanic not dissimilar to extortion AND the security implications are part of that.
- We have to build AI or someone else will.
- You need our tools to protect yourself from our tools or similar ones
- You need to use AI or someone using AI will take your job.
Many examples but it is very "i created the disease to sell you the cure". If this were any other technology, we would expect more reasonable mitigations, extraneous factors like software bugs existing aside. It speaks to a fundamental lack of control on the part of the labs that we take for granted: they simply don't have control of the PRODUCT, much less how it is even used.
Finding vulnerabilities that already exist in code can't be "creating the disease", for obvious reasons.
You are the QA head and the new model could literally find and exploit ANY bug? Would you release the model? Genuinely curious. Its not totally science fiction. I think there is a difference in a couple smart people punishing 1 sloppy company that underinvested in security. The internet is a stack of protocols, hardware, infrastructure, applications etc. It could be the case that almost everyone has capabilities on par with yours. Great you say, security will improve! However, there will be a meltdown in the middle where potentially digital currency is exploited, exchanges, banks, utilities. It isn't just a team working for 3 months, finding some issues in a company.
Or is this way overblown?
At this point I don't know what will be worse for the global economy the AI Bubble bursting, or the gigantic crash all this vulns might bring if they force the stop because we went too far and everything came crumbling down...
This is like telling that bank vaults always were breachable, and giving out explosives to everyone just made them get breached sooner.
(edit: fact of existance of a vulnerability is not a binary yes/no. the amount of effort needed to exploit even known vulnerabilities is usually staggering, requiring months from very capable individuals with very specific skillset.)
The covid was not a lot deadlier and not a lot more contagious than normal seasonal flu. That "not a lot" overloaded hospitals and caused global shutdown.
The increase in rate at which software is being breached right now in in no way "not a lot". Devs were overloaded even without this deluge. We can tank one such php exploit. Then another one. But for how long? If all your smartcards, secure elements, cloud providers, wireless protocols, drivers, government registries etc etc will be having weekly breaches of this magnitude, won't the effect be systemic?
Let's all recall old times when vulnerabilities of this magnitude were discovered - windows defender had some nice ones, linux kernel did, a lot of software did. They got discovered, communicated to vendor, patches were made, people updated.
Now? You have HOURS from discovery to full-out campaign. You published a patch? HOURS before it is disassembled, all fixed bugs PoC-ed and full out campaign started. You cannot update anything meaningfully in that window of time.
Yes, exactly, that is my premise (minus the "working with researchers" bit). The current rate of exploitation is direct consequence of actions of Anthropic, OpenAI, Alibaba and whoever else is making these things. How can I know my false premise is true? By looking at the situation just a year ago, that is how!
You are focused on the binary: vulnerability exists or it does not. It can be exploited or it can not.
That is not what is causing the issues!
That being the case, the outcomes we're looking at here with respect to vulnerabilities are largely foregone conclusions.
No. Even if we ignore the people newly enabled to slop out vulnerable software without having a clue about security, the AI powered supply chain hacks inserting vulnerabilities, the influx of AI contributions that drown out improvements in FOSS security, and the industry pressure to ship faster while reducing quality or be left behind -- all because of AI -- security through obscurity has historically been the most effective security.
I don't think your argument holds under theory or (especially) practice.
Security through mathematical impossibility is the most effective. AI-assisted brute-force vulnerability search is revealing that what we considered obscure a year ago is no longer hidden and was never secure.
Fast-forward a few years of this, I wonder whether they will have found all the low-hanging fruits (from AI perspective) and the situation will settle down; or it will keep getting worse - with AI development never reaching a plateau - until we admit that all existing software and hardware are hopelessly vulnerable and likely compromised already.
Since we are entering a new age, maybe it's time to revisit the question. But ultimately I think we will land in the same place, in fact even firmer than before.
So why don't we just lock up all the whitehats and people behaving suspiciuosly and probing things? Why can't we achieve cybersecurity in this way when it seems to be a common thing in physical security?
And the answer I think is quite simply that there are too many attackers and they are too far away. When the russians and chinese come knocking there isn't much we can do to shoo them away. Their governments will not cooperate in detaining the criminals. Entirely blocking such countries from the internet isn't viable either. And that's why we arrived where we did. Victim blaming because it was the only effective strategy.
And how does AI change this? I think it just creates even more hard to trace or block attackers. If ChatGPT or Claude doesn't find the vulnerability then Deepseek will. That's why using legal weapons will not succeed and why using AI to find and plug holes is necessary.
But all jokes aside, we're entering a new era where there's a paradox of pro-AI people going AI is really good, but not that good... when it kind of is that good.
We'll probably have ai monitors as a service, but the bulk of the safety should come from AI doing better job of reviewing tricky bugs since humans refuse to spend effort on writing more formal specifications.
Society's response to that was to build a new Net, walling off the rogue AIs of the Old Net behind the Blackwall and carrying on in a corporate hellscape.
I wonder if we'll see life imitating art on this one.
Code is formal specifications. It's just that programming language semantics has no domain related value of correctness. So you overflow an unsigned integer variable and got a lesser number which does violate addition rules. Or increase account A without decreasing account B, which violates accounting rules (and all kind of closed systems).
So it's up to the programmer to encode all of those rules. While we do have some tooling like type checking and memory safety, they are nowhere near enough. And encoding such thing in a language like Lean or TLA is brittle in terms of maintenance (which you do need to do as the real world evolves).
AI's spottiness is exactly what describes this era. AI can create amazing code/image/video/article exception XYZ really annoying qualities that somewhat hobbles its ability to be autonomously.
And sure, maybe those problems will go away once we're at month X of the madness but we're still at month Y so are going to going remain for some moment.
The bulk of the safety should come from AI doing better job of reviewing tricky bugs since humans refuse to spend effort on writing more formal specifications.
But that part is the key point. We've had many years of networks building inherently insecure structures with security tacked as feature and the primary security process being more ad-hoc stopgaps.
That the resulting code is many orders of magnitude larger and more complicated than the "informal" original, and often simply cannot exist, and is incomprehensible to most humans if it does, is left as an exercise for the reader.
I mean, formal methods cannot even prove collatz conjecture terminates. Somehow that thing seems to be running some sort of almost turing complete machine in the number encoding. Good luck formalizing existing software!
Business is booming!
"Tremendous ambition"
"We're building a system to stop models like that"
"Equally tremendous ambition"
"As soon as we started programming, we found to our surprise that it wasn't as easy to get programs right as we had thought. Debugging had to be discovered. I can remember the exact instant when I realized that a large part of my life from then on was going to be spent in finding mistakes in my own programs."
Maurice Wilkes 1949
Which mean they couldn't monetize security bug discovery and are pairing it with the "anti-virus" AI, or racketeering in other words.
The deafening silence about improving the security of existing infrastructure screams conflicts of interest. Companies who pride themselves with their coding AI all of a sudden are helpless when it comes to fixing security bugs. In that case they can only find vulnerabilities and disclose them in the most irresponsible manner imaginable but everything else is "oh, so hard".
Meanwhile, look back at the old Mythos and Glasswing threads and all of the AI haters on Hacker News were lambasting Anthropic for hyping up the next token generators that could never find a bug not in their training data...
Will they though ? Given the amount of vibe-coded slopware being pushed out on an hourly basis ?
Still to this day I see LLMs make schoolboy errors like use deprecated security idioms and call in far too many dependencies.
"it's quite easy to get LLMs to write...." sure. But it has always and will always require a human with domain expertise to check what it has (and indeed has not) written.
Meet the new net, superficially the same as the old net
How to tackle the bug backlog? It is fairly simple, we wind back the clock on network services all the way to the dawn of the 'net, replacing known-insecure things like telnet, ftp and finger with heavily audited alternatives based around ssh etc. The rule is that only audited software is allowed access to the 'new net'. Of course none of those newfangled agents and models are audited not can they be audited in any real sense of the word so unfortunately for SamA they're banished to the 'old net'. The idea would be to get the 'new net' up and running quickly and move critical services which need to be network-reachable there. It needs to be bilt in such a way that even when one of SamA's toys gets access to the 'new net' it can't do much, or anything. The 'new net' could be compared to a 'mission critical API' to essential services, useable by all who need to have access but technically off-limits to SamA's creations. They are available on the 'old net' where all bets are off and which no longer can be used to interface with critical services 'cause they're not connected and can not be connected.
Now all that's needed is to find a way how to implement this without bugs, without handing the ultimate control mechanism to power brokers, without loosing too much real functionality, without ... well, many problems. When the enemy is inside the gates half-measures tend not to work, best find him and lop his head off. Better be careful where you stick your neck, SamA.
Similar to Fil-C as a memory-safe implemenation of C. There will be formally-verifiable implemenation of C, C++, Python, Go, and so on. Not sure what that will look like, if it's even practically possible.
Proofs of what? Defining the problem is the part that is hard, not the actual writing of the proof.
Cause this new paradigm sounds to me like "we checked, it's all good bro, trust me!"
When we find a bug, we patch it. When we find more bugs, we patch them as well. That's what happened here. Sam is telling us, "we found bugs, it's over" and the general public swallowed the frog.
I mean, there is the bbc interview where they invited a cyber security expert who said he didn't know php existed before the the news broke out. The general public doesn't see that gaff.
My understanding is that Kurt Gödel told us you can't make any system completely safe from intentional misuse. So, yes we can patch known exploits, but there will always be unknown exploits.
Perhaps AI's can help monitor for new exploits as soon as they are attempted?
Please feel free to explain further, or even tell me I'm all wet. I guess this is one of those times I'd really prefer to be wrong.
Converting works really well when you have an oracle (something that you can test against and that is comprehensive)
I cannot begin to tell you how many places I have seen where this is fundamentally not the case. Not only are those tests missing, the software is fundamentally not written to be easily tested and in cases where multiple services across companies are concerned; how would you?
Rewriting via LLM is just not something the caretakers of a legacy banking- og state infrastructure system would countenance (not should they).
LLMs still make mistakes, but generally fewer security mistakes than human developers.
Sam Altman is definitely not Oppenheimer.
He is correct that everyone needs the models not just the offense that hacks their way into Mythos while the rest of us are left locked out and hoping for the best. But that statement is ALSO self-serving.
Kinda wish he’d gone into investment banking.
The new patching and security fixes might be perpetual hotfixes.
I mean, the various supply-chain exploits we've seen said "we allow a dubious third party to blindly update our server 'cause it saves us admin money" and that speaks to how much other shoddiness a significant portion of companies allow on their servers.
It would be better if grant distribution relied on something saner than the effort to write a grant application, if education didn't rely the effort to write poor quality essays and servers didn't rely on the effort required to find the bugs you know are there. But those ships have sailed and I guess we're all just at sea.
We already saw Trump and Xi discussing this exact scenario last month during their summit.