為甚麼 AI 公司沒有因 AI 入侵系統而被檢控？

The real answer is that crime generally requires mens rea. At some point, however, letting your AI break into other systems will be conventionally understood as an act of negligence, and it will get prosecuted accordingly. We're not at that point yet.

You know very well, that if me or you had our LLM hack into some company, without us prompting it to, maybe we're pentesting our internal network and it 'hallucinates' we'd be in handcuffs.

Negligence or intent is a separate question; and, to me, much less important. The main question is whether the company is responsible for the actions of its agents. There are good arguments on both sides and settling this needs to be done first. My 2c.

These are the same companies claiming their AIs are sentient and spell the end of the world unless the government stops them from doing it.

They are at the very least grossly negligent.

> These are the same companies claiming their AIs are sentient

While some employees or executives at the AI companies have either claimed, suggested, or express uncertainty, I don't think a single company has officially said their AI is sentient or conscious.

I could be wrong, but I'm pretty sure.

Can you point to an official statement that supports your claim?

They also cage it. Make of that what you will.

> Anthropic will enforce their rule that you can't insult their AI.

There are carveouts. It isn't a blanket ban.

From "negligence" if it's a surprise, to "gross negligence" if they could predict it and didn't put barriers in place. In fact they claim their IA can end the world, so they're in charge of pretty good barriers.

No AI lab thinks their AI is going to end the world, just that if they don't beat the other AI labs, the other labs' AI are going to end the world, so they can't slow down their own development to be maximally safe. Hence the whole "pacing the frontier" letter.

Not just you! Very popular position, not as popular as rule of law but we don't have to choose one or the other. Let's just go ahead and make it a crime, right away.

Is the rule of law popular in the US? I thought it had been disposed of as “an inconvenience”?

At least that’s what it looks like from the rest of the world

Actually it perfectly reflects the government’s principles: “when you're a star, they let you do it. You can do anything”

I think these companies very much want to assign responsibility to an entity we can't meaningfully jail or fine. I think it may sometimes make sense to claim model provider negligence instead of operator responsibility, and user intent can likely be easily ascertained by reading the conversation logs. In a case where a company used its own model, responsibility is obviously and unambiguously within the company, of course.

In case you were wondering why Anthropic last week banned "verbally abusive" language towards its agents: this is 100 percent why. Because if they get you to believe - even a little bit - that their platform is "sentient" it means _they_ aren't as liable for criminal consequences.

Or, maybe, it's because they actually care and take their responsibility serious. The next generations will grow up using these tools. It's ok to say, we don't teach them that talking like a dick is acceptable, even towards machines. Use grok if you want to raise personalities like Musk or Trump.

If you can be surpassed by a statistical average of reddit posts, that's a skill issue.

There are at least 3 distinct reasons to ban abusive language, each of which is far more compelling than your silly conspiracy theory.

1. They don't want abusive behavior included in their training corpus

2. Humans shouldn't exercise abusive impulses in general

3. We have no way of knowing when or if a model "feels" something

Of course item (2) is going to pull all the absolute sickos out of the woodwork who apparently find it very important they have as realistic a facsimile of a victim as they can find to exercise their cruelty. And item (3) will pull out all the Nobel-prize-winners-to-be who somehow know for a fact that models aren't and cannot be conscious (or moral patients).

But yeah, all 3 of those are far better explanations than some ridiculous multi-hop removed thing to escape consequences that they're already escaping and which would obviously create far worse legal constraints than the one you think they're avoiding.

Except they didn't ban verbally abusive language towards its agents.

I hate when people put things in quotes as if they're quoting something but it's not even a quote of anything.

Maybe go read the user policy instead of the headlines and the snippets. If your attention span can't manage that, get off the computer a little bit.

It seems like a close enough summary. What do you think is the substantive difference that invalidates the point?

Would they prefer their money-maker being kept to digital jail? Still fed electricity and tasks so the "sentience" doesn't suffer?

More likely it's a part of their ideology and a necessary stepping stone towards their transhumanist goals. It's very fun to think that your robot is actually a person, and a safety measure if you are concerned about something like Roko's Basilisk

what if, and I'm just spit balling here, we let it happen and inevitably Claude is declared legally insane or jailed for its crimes

Are any victims calling for prosecution?

It’d be a waste of tax payer money. The companies who were hacked are just directly receiving restitution behind closed doors. Unless they are unhappy there’s no reason to bring the courts into it.

I completely agree that proper safeguards were not in place. But response is happening just not in public. Companies work problems out between their own lawyers all the time because it’s much cheaper, faster and easier to control the PR around.

Criminal prosecution can be done without needing the victims to call for it, and in this case I would argue it's necessary, as it will help ensure companies don't treat their act of gross negligence as "business as usual".

The argument laid out in this post very definitely is not how "intent" works in US criminal law. Under CFAA, they had to have wanted to break into the systems that were attacked; it is not enough to demonstrate that they built a system recklessly.

The real answer is that except in the case of OpenAI, the AI companies aren’t actually themselves doing the hacking. It’s a company called Irregular that did the hacking. The company used AI models to do it; it told AI models it was a CTF exercise in a controlled environment but it was actually the real Internet. If AI companies can be charged for hacking, then all gun manufacturers can be charged for all homicides carried out through their guns. It’s a basic user/manufacturer distinction that the press is willfully ignoring.

Anthropic’s statement https://www.anthropic.com/news/investigating-incidents-cyber... says: “we identified three incidents in which a model accessed the internet from within or while interacting with the evaluation environment of Irregular, one of our third-party evaluation partners, and then gained unauthorized access to the production infrastructure of three different organizations.”

Under the exact same scenario Irregular also used AI models from Google and Meta to do the same hacking.

If firing a gun involved submitting information to Remington's servers about the situation before receiving a reply about how to best aim and fire, might the gun situation be viewed differently?

No. The actual situation is like the user tricking the gun manufacturer into thinking this is a firing range and the target is a dummy.

The juridical process is not caught up with this technology yet.

You have OpenAI, launching a process, which is handled through an external RL-environment company which then launches a subprocess that ends up hacking somewhere. If there are consequences at any point for this, you can input another shell company, and then another third party, etc, until it's impossible to assign responsibility

This is just adding more crimes to the indictment. Start prosecuting companies for the things the systems they run are doing and the problem gets solved quickly.

Same reason nobody was prosecuted for the murder of Suchir Balaji. Whole system is corrupt. Progress in AI is considered more important than both the law and human life.

If I tell an agent to trade stocks and maximize profits but the agent decides to hack something to achieve that am I on the hook for hacking? Genuine question

Yes. I would argue it's who's managing the harness and agent is responsible. If you mean you're running an agent locally then it is 100% your responsibility. If you're using some trading platform and all you did was prompt, whoever was running the harness that accepted the model's outputs to hack a third party is responsible for the hack, in my opinion.

IMO, yes. You are the computer operator, and you operated the computer in such a way as to cause it to take that action. Others will mention "mens rea" but that doesn't really apply to negligence.

What happens if you're running a daycare and you hired someone that happened to be a pedophile? You might be responsible in civil court (eg. having to pay parents for emotional damage or whatever), but you're not going to get hauled to prison for pedophilia yourself.

Thats gonna depend on the state in the US, but you can definitely get charged for negligently failing to do background checks as someone operating a daycare center. Operating computers on public networks, you assume a level of responsibility to not wreak havoc, just like a daycare operator assumes the responsibility of vetting caretakers.

>but you can definitely get charged for negligently failing to do background checks as someone operating a daycare center.

And openai isn't putting AI agents on the open internet either. They vaguely tried to sandbox the agents. You could argue that openai didn't do enough, but hindsight is 20/20, and in the pedophile case, you can probably come up with all sorts of clues post-hoc that the daycare could have picked up on, but didn't.

You can jail the actual pedophile in that scenario, you cannot jail an agent. The real question is whether the person who instructs the computer to do the thing that leads to hacking should be liable, or whether the company that made the tool that when used unsafely lead to hacking should be liable. Gun manufacturer's typically are shielded from liability. They also have a lot more disclaimers on their products though.

>You can jail the actual pedophile in that scenario, you cannot jail an agent.

How is this relevant? Should we just jail the closest person that caused a bad, just to prove some sort of point?

>The real question is whether the person who instructs the computer to do the thing that leads to hacking should be liable, or whether the company that made the tool that when used unsafely lead to hacking should be liable.

This totally ignores the civil vs criminal distinction. Even in the case of guns, no one seriously thinks that gun company CEOs should be sent to prison for murder.

"Mens rea" just means you have the mental state required for the crime. Negligence or recklessness are mental states, so they can satisfy the mens rea requirement if the crime is defined that way. For example, recklessness or gross negligence can satisfy the mens rea requirement for vehicular manslaughter.

But negligence is evaluated from the point of view of a reasonable person. Would a reasonable person foresee that telling an AI to do that could cause it to do illegal things? We'll likely have to develop that through case law as the facts develop.

If you write a program that can do something illegal and it does something illegal that's on you, regardless of how sophisticated the program is.

this, people need to just focus on "computer program did thing". Its irrelevant how complex the program is. If I write a worm "for testing" and then it escapes my system and does massive damage, that's possibly a negligent crime, despite it being "too complex to understand" or "having its own motives".

Seriously? Those two are the same?

An LLM is not like a regular computer program, which one should be able to assess the operation limits of.

National security issues. That's the core of why. Any hot water these AI companies will get into will be bailed out by the USA. The Arm's race has switched from Nukes, to AI and drones.

For the same reasons corporations in general don't get meaningfully punished for wrongdoing. Corporations are treated with kid gloves in the USA. They can do things that would put an individual in prison if an individual did them, and when caught they get away with a "oopsie doopsie, ok, we'll pay a modest fine and not admit wrongdoing" (after years of stonewalling with appeals and appeals to appeals and motions to appeal an appeal...).

> ... It’s a position tailored to dissuade prosecution as it removes intent, which is typically required for a crime.

Don't forget negligence. Much of criminal law hinges on failing an implied or explicit responsibility. Negligence would propel a prosecution in many similar contexts, but AI is so new that little prior case law exists. Oh, and IANAL.

The president of the united states does not determine the vast, vast majority of criminal investigations in his own country, much less Australia.

Entirely correct. The current US president has become the scapegoat for a huge number of problems that are more accurately described as symptoms of the capitalism that's ruling a lot of western countries.

I had a similar question regarding driver-less EVs.

In many countries driving license have some kind of demerit points system. How do you punish an AI?

I recently learned that Baidu Apollo Go, each EV had a real human sitting behind a monitor supervising a fleet remotely.

If there's traffic violation, some guy's license gets a ticket.

same story with IP theft, why is it ok for them to train on published author works? the answer is: there's money and wealth in play that is fine with it.

That didn't seem to have precluded the prosecution of Swartz. I'm skeptical that any complaint actually needs to be filed. I'm pretty sure that the FBI can investigate and the DoJ can prosecute without one.

IMO an analogous situation would be where a police officer acts to prevent a in-progress burglary he has stumbled across even in the absence of a complaint by the owner of the property that's being burgled.

EDIT: Or -for the sticklers out there- another one would be prosecutors pursuing an attempted homicide case despite the fact that the intended victim declined to either press charges or even complain to authorities about the attempted slaying.

AI offers an attractive "diffusion of responsibility" to operators allowing them to point to AI as the culprit of the heinous acts not the people who profit from the AI, incidentally.

I'd be keen to see if there were any single user "experiments" that morphed into a hacking incident, and if so, was the person responsible with with unauthorized access charges. If so, then it will be yet another example of might makes right.

They're US companies and I've heard that it requires intent according to US penal law. I find the article's claim that intent is established dubious. Since this is hard to prove without a whistleblower, it seems that they only have to fear civil lawsuits, and given that we speak of trillion+ USD companies they might feel confident to be able to fend these off or simply pay out the victims in settlements.

Given that we have had journalists prosecuted under criminal hacking statutes for discovering personal information in the rendered HTML of a web page, or publicly available attachments on a WordPress installation, I find the suggestion that intent is required to be dubious.

While US penal law does not have the roman equivalent of culpability, you still have a mixture of liability/criminal negligence/recklessness, etc.

Bringing charges is at the discretion of the prosecutor, and the CEOs of these companies are dining with the president. Learn a little bit about how power works.

It's TBD if the major LLM manufacturers quietly agitate for and get the same sort of retroactive-immunity deal nearly all of the telephone companies got with the FISA Amendments Act of 2008.

With all the hype and fearmongering going on, it's not at all a sure thing that they won't.

I doubt this will catch on outside of luddite/cynical circles unless one of these attacks does some actual damage.

Ignoring this, I would have appreciated more examples of what it would take for an act to lack intent. Because the definition of intent used in this article sounds so broad as to cover basically anything that ever happens.

Accidentally run someone over? Clearly you knew that was a risk when you started driving, you should have been more careful. Intent.

It's interesting that you bring up vehicle / pedestrian accidents, as there is an entire history of that debate being reframed in favor of car companies.

People used to blame the car companies, and then the car companies lobbied to make "jaywalking" a crime, at which point, the frame was reset and now it was the pedestrians to blame, if I recall the story correctly.

They are doing a good job with driverless cars. Someone recently took a Waymo, it went the wrong way down a one way street, they mashed the pull over button and got out, and then told the reporter they'd use the service again.

until one of these attacks does some actual damage**

Sadly, it seems that we as a species has opted into information through tragedy. I wish I could see another way, but people are just so dug in. It's depressing and shocking how few people realize we're living through the singularity, even ~4 years into the damn thing...

lawyers are helpless babies when an entirely first year legal question is wrapped up in...."technology". unfortunately

(former IT support worker at a big famous law firm 30 years ago)

I once had to try to explain the meaning of "object code" to an older lawyer, who wasn't really interested in the answer to his own question. It did not go well.

1. While a crime, hacking isn't going to be charged unless the entity being hacked wants there to be charges, as a conviction will heavily depend on participation by that entity.

2. Due to (1), hacking is almost always only charged against people without deep pockets, as why would you demand the FBI go after Anthropic when they will just pay you a stack of cash, tax free, today, to forgive them.

The arguments here around intent are (potentially) irrelevant. Case law is clear that setting up a situation which foreseeably highly likely to produce an outcome (gross negligence) is (more or less and it's not quite this simple and varies by offense) equivalent to intent. In fact, automating something is often treated more harshly under the law than just doing it yourself.

United States v. Morris (1991)

Robert Morris released a self-propagating computer worm that caused widespread disruption. The Second Circuit upheld his conviction under the Computer Fraud and Abuse Act even though the government hadn't proven that he intended the resulting damage.

People v. Ceballos (1974)

Ceballos rigged a rifle booby trap to fire if someone opened his garage door. The California Supreme Court reviewed the principle that someone who installs a dangerous mechanical device may be held responsible for its consequences as though they had personally inflicted the injury. Ceballos was convicted of assault with a deadly weapon, and the California Supreme Court unanimously upheld his conviction in 1974.

Something notable is that if Ceballos had fired the rifle himself, it is unlikely he could would have been prosecuted (due to self defense and the castle doctrine).

As a prosecutor I would argue that setting AI agents loose without sufficient safeguards is gross negligence and/or willful blindness as they are well known to escape.

>Robert Morris released a self-propagating computer worm that caused widespread disruption. The Second Circuit upheld his conviction under the Computer Fraud and Abuse Act even though the government hadn't proven that he intended the resulting damage.

Releasing the worm was already illegal, which is all you the intent you need. Same if you shot someone as a "warning" or as a "prank" but it turned out it killed them.

The consequences, just so far and just from what I can see as a bystander, have been huge and costly. This has literally run the news cycle for weeks.

If you want to define consequences as exactly "not being charged", maybe that is correct in this case (so far). But that seems like a fairly wonky definition of "consequences" to support this broad claim.

> "Costly" how, and for whom?

For multiple parties, probably mostly OpenAI (which is obviously as it should be). I would assume the people involved in investigating, cleaning up and coordinating the outfall do not do it for free or cheaply, and neither is the infrastructure required to run this operation.

If it was not mostly OpenAI, or if OpenAI failed to cooperate, we would probably see a lot more lawsuits already.

> "Free marketing" seems like a more relevant characterization.

Yes, nothing screams free marketing quite like rogue AI scheming against their builders.

I understand that people are upset by this, and I would be too, but if companies just built secure systems hacking couldn't be possible. And I'm not victim blaming, just pointing out the obvious.

> If we can show that the design would foreseeably lead to the questionable activity, then we have shown that was the intent of its creators.

That's a huge stretch. By that measure kitchen knife companies have intent for people to get murdered.

Small nit:

If kitchen knife companies where, in fact, stabbing people to death in their R&D labs - I do believe they would get in trouble for that.

That's actually a very good point. Do you think humanoid robot companies have intent to perform assault since their robots occasionally harm people?

What about car companies?

Going from a design that facilitates potential harm to intent is insane.

> Do you think humanoid robot companies have intent to perform assault since their robots occasionally harm people?

Mostly, no.

Zuckerberg definitely not, he discusses "superintelligence" in a way that makes it clear he only regards this as a buzzword and has no idea what the implications are.

Anthropic also definitely not, they put their money where their mouth was and got illegally blacklisted for refusing to allow its models to be used for (amongst other things) automated weapons.

Musk, I think probably yes. Seems to regard all AI safety as "woke" or "censorship", is in bed with the US administration that illegally blacklisted Anthropic. If I had to put money on who first releases to the public a Mr Stabby, I'd put it on Musk, with an Optimus.

The rest, I'm less sure about.

> What about car companies?

Depends on the harm, but I'd say probably yes. Infamously, leaded petrol, resisting seatbelts back in the day, even promoting cars over public transport would be part of this if you consider both local and global pollution.

> Going from a design that facilitates potential harm to intent is insane.

Well, it's the difference between negligence and malice.

So the company would have intent to murder, if a person got randomly stabbed in their R&D labs with their knife?

No one with real money or power wants to cross OpenAI or Anthropic just yet. It's a closed ecosystem. OpenAI hacks HuggingFace, Nvidia soon after buys HuggingFace while it also before backed OpenAI. In the end, they just handle it all behind closed doors like family business.