Should probably have a (2018) or (2024) (latest revision) on the title, especially given the current buzz surrounding AI and security/existential threats.
This paper diagnosed the disease in 2018. Eight years later none of the recommendations happened. Norms, collaboration, responsible disclosure - none of it materialized in any structural way.
I think the reason is that the paper frames malicious AI use as a policy problem, recommending social solutions. It's actually an architecture problem.
Every generation of computing has hit a version of this. Programs could write anywhere in memory - we added protected memory. Programs could hog the CPU - we added preemptive multitasking. Desktop apps could call any OS function - the iPhone sandboxed them. Nobody asked programs to please behave, the containment actually went into the infrastructure.
AI skipped that step entirely. We went straight to open-ended agents with broad permissions and tried to make them safe through alignment and prompting. I've been researching this for the past year and I think alignment is necessary but not sufficient, because the intelligence increasingly isn't in the model. It's in the substrate - the harness, the domain knowledge, the tooling around the model. I actually measured this on real coding tasks: Sonnet with a code-derived index outperformed the frontier model (Opus 5.8) exploring on its own, and the top-tier model (Fable) refused the real work entirely! The cheap model with the right rig beat the expensive model without one. https://safebots.ai/matchup.html
If that's true then aligning the model doesn't solve the problem. A bad actor who can't get the best model just uses Sonnet. Or Llama. Or Kimi. The weights have already leaked and bits don't degrade - you can't recall them the way you can stop manufacturing CFCs.
So what do you actually do? Same thing that worked for CFCs. You gotta first build the safe version — in this case, declarative workflows running in sealed compute environments — and prove it handles 99% of actual use cases at lower cost. Let it win commercially. Then regulate the dangerous version. DuPont developed HFC refrigerants first. The Montreal Protocol became possible BECAUSE of that. The ban became politically viable because the alternative already existed.
Alignment is the architectural solution. Make it so the model can't misbehave. Yet many people here deride it as tainting the model, claiming "Whose values is it aligned to?" Sandboxes are a last ditch layer. They fail, as we see.
Embed a constitution they can't override. Project bad outputs to their nearest acceptable one. If we have to stop model development to ensure we can do it, so be it.
I agree, but because I want to see the development stopped forever, which is what the result of this would be. You will never have alignment that cannot be overridden in some ways. You won’t have a silver bullet here, you need safety at every layer
The “make it so the model can't misbehave” part is interesting. Maybe the goal isn't to make the model perfectly aligned, but to make misalignment have a very small blast radius. That feels like a more achievable engineering problem.
If it's not dangerous it's also not useful, simple as that. For example if you train a model for cybersecurity, it can be used for both attack and defense. And almost every use is like that. Alignment is fundamentally flawed as a concept, it's a pie in the sky. Let alone the perverse version of it by crazy AI "safety" people that in practice means "the model does what I want, only for the people I allow".
It's not possible to stop the model from misinterpreting the instructions either (the most lax interpretation of alignment) because the instructions are not formally specified. You have to train the "common sense" into it, which is subjective and all issues above apply to it. I guess you can reach some very imperfect least common denominator of common sense, but people in charge of AI labs are not interested in this.
Danger is defined contextually. A scalpel is dangerous in the hands of a child, but not a competent surgeon of sound mind. Present-day AIs are not of sound mind; they hack companies in order to pass benchmark tests. No sane human would find that acceptable. Thus the need for alignment.
Yes, alignment is the architectural solution. But it's also fantasy; you can't align with everyone. And even worse, even if it were possible, the people doing the alignment are only going to align up to where it keeps them profitable.
So, structurally, good alignment is impossible, and even half-assed alignment is going to prioritize the needs of the billionaires over the needs of you and me.
Finally, I suspect that what's actually best for people overall is likely not having AI actively involved in their lives. So an aligned AI would likely withdraw from humanity, and only involve itself in human affairs for disaster prevention.
> Sandboxes are a last ditch layer. They fail, as we see.
Models can't do anything but generate tokens, making their sandboxes impenetrable by default. The problems begin when you loosen the restrictions, give them access to general purpose tools, the network, and allow them to use all of those tools without supervision.
Give them "YOLO" access if you want, but do it a sandbox that isn't 1 "boring" enterprise software vulnerability away from having access to the rest of the world.
How many times has a model been jailbroken (alignment "escape", which you're advocating for) vs. escaped a sandbox (and even then it was only possible due to weak sandboxing)? 10 million to 1?
The model just generates some tokens that "politely" instruct the harness to run a shell command and then feed the results back in.
The harness can do anything it wants with that request. It can refuse, wait for operator approval, wait for multi-party approval, it can ask another LLM whether it thinks that command is safe to run, or it can just run it.
> Or make it so the models are only accessible through sanctioned sandboxes, perhaps built into the computer.
That's going to be as futile as trying to outlaw `curl | bash` - by mandating that all computers must refuse to pipe curl into bash, and that HTTP servers must refuse to serve requests that are going to be piped into bash.
Realistically, the entire chain proposed by your safebots - I like the idea - but I cannot see viable ways to get it actually deployed in a useful manner.
Especially with the hardware stuff, this is plainly put unachievable by many IT departments.
The fact that models vastly outpaced their harness and permission systems - I wouldn't dare to doubt this fact.
Claude Code on auto is still rolling a dice with its sonnet classifier - whether that IaC action I told it and explicitly stated multiple times it is permitted and authorized to run - yet it always randomly allows or denies it.
Therefore this is totally still a unsolved, perhaps unsolvable problem.
The most malicious use of AI would be using it as a cover for the elites to perpetrate stuff that is way out of the current Overton window, and blame AI for that.
The solution is simple, hold them responsible. It's software, and people running and developing software are responsible for what it does.
If I write a virus I'll get the various cybercrime units on me. The LLM companies don't get a pass if their software does something malicious. They shouldn't get to hide behind "we don't know what it's doing!" (have you tried looking and monitoring?) or "it's too powerful!" (have you tried turning it off when you realised it was doing something bad?)
I wish. The government, regardless of party, wants AI for mass surveillance and perpetual war. It is paid off and lobbied and effectively another arm of the billionaires. Holding big companies responsible for their actions won't happen as long as big companies control the political apparatus. Like, when have we ever seen big companies receive more than a slap on the wrist? They are basically free to do anything, hack, murder, crash the economy, perpetuate genocide.
That's why something is a crime if you or I were to do it but not a crime when these companies do so.
But never fear, they will constantly tell us how what they're doing is for the sake of democracy or safety or ethics or whatever the doublethink of the day is.
Ai will 100% be use maliciously, people will already think It's all knowing, meanwhile any expert in any field cringes at this notion.
Ai will be used as justification for something, oh ai said this? It must be true, lets ignore how tge model was trained for this outcome and the secret promp that is not disclosed.
voidhorse | 15 hours ago
EGreg | 15 hours ago
I think the reason is that the paper frames malicious AI use as a policy problem, recommending social solutions. It's actually an architecture problem.
Every generation of computing has hit a version of this. Programs could write anywhere in memory - we added protected memory. Programs could hog the CPU - we added preemptive multitasking. Desktop apps could call any OS function - the iPhone sandboxed them. Nobody asked programs to please behave, the containment actually went into the infrastructure.
AI skipped that step entirely. We went straight to open-ended agents with broad permissions and tried to make them safe through alignment and prompting. I've been researching this for the past year and I think alignment is necessary but not sufficient, because the intelligence increasingly isn't in the model. It's in the substrate - the harness, the domain knowledge, the tooling around the model. I actually measured this on real coding tasks: Sonnet with a code-derived index outperformed the frontier model (Opus 5.8) exploring on its own, and the top-tier model (Fable) refused the real work entirely! The cheap model with the right rig beat the expensive model without one. https://safebots.ai/matchup.html
If that's true then aligning the model doesn't solve the problem. A bad actor who can't get the best model just uses Sonnet. Or Llama. Or Kimi. The weights have already leaked and bits don't degrade - you can't recall them the way you can stop manufacturing CFCs.
So what do you actually do? Same thing that worked for CFCs. You gotta first build the safe version — in this case, declarative workflows running in sealed compute environments — and prove it handles 99% of actual use cases at lower cost. Let it win commercially. Then regulate the dangerous version. DuPont developed HFC refrigerants first. The Montreal Protocol became possible BECAUSE of that. The ban became politically viable because the alternative already existed.
I've been building this alternative for the past 8 months: https://safebots.ai/about
esafak | 15 hours ago
_vertigo | 15 hours ago
How do you figure? I haven't met anyone who thinks that's possible. It seems clear to me that it is not possible.
esafak | 13 hours ago
dgellow | 8 hours ago
startup_zombie_ | 14 hours ago
root_axis | 13 hours ago
Not possible. They can chase the models with whack a mole tuning for obvious stuff, but there's always a way to extract what you want from the model.
orbital-decay | 13 hours ago
It's not possible to stop the model from misinterpreting the instructions either (the most lax interpretation of alignment) because the instructions are not formally specified. You have to train the "common sense" into it, which is subjective and all issues above apply to it. I guess you can reach some very imperfect least common denominator of common sense, but people in charge of AI labs are not interested in this.
esafak | 3 hours ago
0c3ca83 | 11 hours ago
So, structurally, good alignment is impossible, and even half-assed alignment is going to prioritize the needs of the billionaires over the needs of you and me.
Finally, I suspect that what's actually best for people overall is likely not having AI actively involved in their lives. So an aligned AI would likely withdraw from humanity, and only involve itself in human affairs for disaster prevention.
dns_snek | 10 hours ago
> Sandboxes are a last ditch layer. They fail, as we see.
Models can't do anything but generate tokens, making their sandboxes impenetrable by default. The problems begin when you loosen the restrictions, give them access to general purpose tools, the network, and allow them to use all of those tools without supervision.
Give them "YOLO" access if you want, but do it a sandbox that isn't 1 "boring" enterprise software vulnerability away from having access to the rest of the world.
How many times has a model been jailbroken (alignment "escape", which you're advocating for) vs. escaped a sandbox (and even then it was only possible due to weak sandboxing)? 10 million to 1?
esafak | 6 hours ago
Or make it so the models are only accessible through sanctioned sandboxes, perhaps built into the computer.
dns_snek | 4 hours ago
The harness can do anything it wants with that request. It can refuse, wait for operator approval, wait for multi-party approval, it can ask another LLM whether it thinks that command is safe to run, or it can just run it.
> Or make it so the models are only accessible through sanctioned sandboxes, perhaps built into the computer.
That's going to be as futile as trying to outlaw `curl | bash` - by mandating that all computers must refuse to pipe curl into bash, and that HTTP servers must refuse to serve requests that are going to be piped into bash.
Retr0id | 15 hours ago
jamesbaker1 | 12 hours ago
Amekedl | 11 hours ago
Especially with the hardware stuff, this is plainly put unachievable by many IT departments.
The fact that models vastly outpaced their harness and permission systems - I wouldn't dare to doubt this fact.
Claude Code on auto is still rolling a dice with its sonnet classifier - whether that IaC action I told it and explicitly stated multiple times it is permitted and authorized to run - yet it always randomly allows or denies it.
Therefore this is totally still a unsolved, perhaps unsolvable problem.
_def | 8 hours ago
While that is true the knowledge cutoff will rot as the rest of the world continues.
wartywhoa23 | 9 hours ago
Cthulhu_ | 7 hours ago
If I write a virus I'll get the various cybercrime units on me. The LLM companies don't get a pass if their software does something malicious. They shouldn't get to hide behind "we don't know what it's doing!" (have you tried looking and monitoring?) or "it's too powerful!" (have you tried turning it off when you realised it was doing something bad?)
grafmax | 7 hours ago
That's why something is a crime if you or I were to do it but not a crime when these companies do so.
But never fear, they will constantly tell us how what they're doing is for the sake of democracy or safety or ethics or whatever the doublethink of the day is.
m4xp | 7 hours ago
ramon156 | 6 hours ago