> You might even convince yourself that these questions are “privacy preserving,” since no human police officer would ever rummage through your papers, and law enforcement would only learn the answer if you were (probably) doing something illegal.
Something I've started to see happen but never mentioned is the effect automated detection has on systems: As detection becomes more automated (previously authored algorithms, now with large AI models), there's less cash available for individual case workers, and more trust at the managerial level on automatic detection. This leads to false positives turning into major frustrations since it's hard to get in touch with a person to resolve the issue. When dealing with businesses it's frustrating, but as these get more used in law enforcement, this could be life ruining.
For instance - I got flagged as illegal reviews on Amazon years ago and spent months trying to make my case to a human. Every year or so I try to raise the issue again to leave reviews, but it gets nowhere. Imagine this happening for a serious criminal issue, with the years long back log on some courts, this could ruin someones life.
More automatic detection can work (and honestly, it's inevitable) but it's got to acknowledge that false positives will happen and allocate enough people to resolve those issues. As it stands right now, these detection systems get built and immediately human case workers get laid off, there's this assumption that detection systems REPLACE humans, but it should be that they augment and focus human case workers so you can do more with less - the human aspect needs to be included in the budgeting.
But the incentives aren't there, and the people making the decisions aren't the ones working the actual cases so they aren't confronted with the problem. For them, the question is why save $1m when you could save $2m? With large AI models making it easier and more effective to build automated detection I expect this problem to get significantly worse over the next years.
Update this to a world where every corner of your life is controlled by a platform monopoly that doesn't even provide the most bare-bones customer service and yeah, this is going to get a lot worse before it gets better.
2. But phones can send plaintext back to the cloud to get help doing AI things
3. And we tend not to know because it’s all “assisstance”
But the solution like anything is pricing. I mean yet again (uber, Airbnb) billions of dollars of VC money is used as subsidy so my photos can get OCR’d.
If phones said “hey for a dollar fifty I can work out what the road sign says behind your dogs head in 32 photos your mum sent you last week” I think we woukd see a different threat landscape
This is - again - unsustainable cash spending distorting markets and common sense. If the market was “we can OCR and analyse these insurance claims” the. Things like privacy and encryption would be first class requirements and harder to sell and build.
By spending a billion they can sell services to people without regulators to ask awkward questions and then they hope step 3. Profit.
I short not even AI can spot patterns in encrypted data, it’s only when plaintext gets sent around in the hope of profit do we see a threat. That’s seems a simple fix if not an easy one
> Yet this approach is obviously much better than what’s being done at companies like OpenAI, where the data is processed by servers that employees (presumably) can log into and access.
No need for presumption here: OpenAI is quite transparent about the fact that they retain data for 30 days and have employees and third-party contractors look at it.
> Our access to API business data stored on our systems is limited to (1) authorized employees that require access for engineering support, investigating potential platform abuse, and legal compliance and (2) specialized third-party contractors who are bound by confidentiality and security obligations, solely to review for abuse and misuse.
> We are about to face many hard questions about these systems, including some difficult questions about whether they will actually be working for us at all.
And how. I'd lean towards no. Where we're headed feels like XKEYSCORE on steroids. I'd love to take the positive, optimistic bent on this, but when you look at where we've been combined with the behavior of the people in charge of these systems (to be clear, not the researchers or engineers, but c-suite), hope of a neutral, privacy-first future seems limited.
My guess is that the main purpose of agents will be to train the AI on your data. Companies have run out of data on the internet for training AI's, so they'll use agents as an excuse to get access to your personal real-time data. This has always been the business model, you are the product.
Given how politics and companies evolved, I actually trust those people in charge of XKEYSCORE systems more than ever. They may wear suits, but those people usually come from some military background, and have a sense of duty towards defending US, from threats both foreign and domestic, and historically have not really abused their powers no matter what the administration is. XKEYSCORE for example, wasn't really about hacking people, it was just about collecting mass metadata and building profiles, well within the legal system, and the blame should be on the companies that didn't provide privacy tools, because any big government could have build the same system.
Meanwhile, the anti anti-establishment Republican Party since 2016 who cried about big tech turned out to be the biggest pro-establishment fans, giving Elmo an office in a white house and Zucc bending a knee to avoid prosecution.
With these new systems, Id rather have smart people who only work in US defensive forces because of a sense of duty (considering they could get paid much more in the private sector) in charge.
It's not a search if we don't find anything, and it's not a seizure if we charge the money with the crime. These are court approved arguments, so they must be correct interpretations.
Point is: modern bureaucrats have proven that they are absolutely willing to abuse power, even in the best of times when there is no real domestic political strife.
The most depressing realization in all of this is that the vast treasure trove of data that we used to have in the cloud thinking it was not scannable even for criminal activity has now become a vector where we shall have thought police coming down upon us for simple ideas of dissent.
A lot of people tried to sound the alarm. It's not "the cloud", it's "other people's computers". And given that other people own these machines, their interests - whether commercial or ideological - will always come first.
To be fair, most people understand that risk. It’s just it is very convenient in a lot of scenarios and some businesses might have not even started without it. Privacy is not that big of a concern for a big chunk of people. And they’re basically voting with their wallets.
> Apple even says it will publish its software images (though unfortunately not the source code) so that security researchers can check them over for bugs.
I think Apple recently changed their stance on this. Now, they say that "source code for certain security-critical PCC components are available under a limited-use license." Of course, would have loved it if the whole thing was open source. ;)
> The goal of this system is to make it hard for both attackers and Apple employees to exfiltrate data from these devices.
I think Apple is claiming more than that. They are saying 1/ they don't keep any user data (data only gets processed during inference), 2/ no privileged runtime access, so their support engineers can't see user data, and 3/ they make binaries and parts of the source code available to security researchers to validate 1/ and 2/.
Note: Not affiliated with Apple. We read through the PCC security guide to see what an equivalent solution would look like in open source. If anyone is interested in this topic, please hit me up at ozgun @ ubicloud . com.
Yes. I made that point a few weeks ago. The legal concept of principal and agent applies.
Running all content through an AI in the cloud to check for crimethink[1] is becoming a reality. Currently proposed:
- "Child Sexual Abuse Material", which is a growing category that now includes AI-generated images in the US and may soon extend to Japanese animation.
- Threats against important individuals. This may be extended to include what used to be considered political speech in the US.
- Threats against the government. Already illegal in many countries. Bear in mind that Trump likes to accuse people of "treason" for things other than making war against the United States.
- "Grooming" of minors, which is vague enough to cover most interactions.
- Discussing drugs, sex, guns, gay activity, etc. Variously prohibited in some countries.
- Organizing protests or labor unions. Prohibited in China and already searched for.
Note that talking around the issue or jargon won't evade censorship. LLMs can deal with that. Run some ebonics or leetspeak through an LLM and ask it to translate it to standard English. Translation will succeed. The LLM has probably seen more of that dialect than most people.
"If you want a vision of the future, imagine a boot stepping on a face, forever" - Orwell
A cynic in me is amused at the yet unknown corporation being placed under investigation due to a trigger phrase in one of the meetings transcribed incorrectly.
The author helpfully emphasized the interesting question at the end
> This future worries me because it doesn’t really matter what technical choices we make around privacy. It does not matter if your model is running locally, or if it uses trusted cloud hardware — once a sufficiently-powerful general-purpose agent has been deployed on your phone, the only question that remains is who is given access to talk to it. Will it be only you? Or will we prioritize the government’s interest in monitoring its citizens over various fuddy-duddy notions of individual privacy.
I do think there are interesting policy questions there. I mean it could hypothetically be mandated that the government must be given access to the agent (in the sense that we and these companies exist in jurisdictions that can pass arbitrary laws; let’s skip the boring and locale specific discussion of whether you think your local government would pass such a law).
But, on a technical level—it seems like it ought to be possible to run an agent locally, on a system with full disk encryption, and not allow anyone who doesn’t have access to the system to talk with it, right? So on a technical level I don’t see how this is any different from where we were previously. I mean you could also run a bunch of regex’s from the 80’s to find whether or not somebody has, whatever, communist pamphlets on their computers.
There’s always been a question of whether the government should be able to demand access to your computer. I guess it is good to keep in mind that if they are demanding access to an AI agent that ran on your computer, they are basically asking for a lossy record of your entire hard drive.
> The author helpfully emphasized the interesting question at the end
We're already there. AI or not doesn't affect the fact that smartphones gather, store, and transmit a great deal of information about their users and their users' actions and interests.
> (in the sense that we and these companies exist in jurisdictions that can pass arbitrary laws; let’s skip the boring and locale specific discussion of whether you think your local government would pass such a law)
Anyway the idea of what’s a reasonable search in the US has been whittled away to almost nothing, right? “The dog smelled weed on your hard drive.” - A cop, probably.
See also CrypTen, Meta's library for privacy preserving machine learning: https://github.com/facebookresearch/CrypTen. This isn't fully homomorphic encryption, but it is multi-party computation (MPC), which hides the inputs from the company owning the model.
But while not revealing user input, it would still reveal the outputs of the model to the company. And yeah, as the article mentions, unfortunately this kind of thing (MPC or fully-homomorphic encryption) probably won't be feasible for the most powerful ML models.
Maybe a little off topic, but is there a way for a distributed app to connect to one of the LLM companies (OpenAI, etc.) without the unencrypted data hitting an in-between proxy server?
An app I'm building uses LLMs to process messages. I don’t want the unencrypted message to hit my server - and ideally I wouldn’t have the ability to decrypt it. But I can’t communicate directly from client -> LLM Service without leaking the API key.
This provides a way for your server to create a limited-time API key for a user which their browser can then use to talk to OpenAI's API directly without proxying through you.
I love this idea, but I want it for way more than just the WebRTC API, and I'd like it for other API providers too.
My ideal version would be a way to create an ephemeral API key that's only allowed to talk to a specific model with a specific pre-baked system prompt (and maybe tool configuration and suchlike) and that only works for a limited time and has a limited token budget.
Will such processing be cheap enough to be done by a box that plugs into a customers router to handle such? Would they buy them? Notably not just for this use case but others
There is always going to be a gap between local-first processing and what can be achieved in a full-sized datacenter/cloud. That leads to the risks mentioned in the article.
I wrote about Apple's Private Cloud Compute last year; for the foreseeable future, I still think server-side Confidential Computing is the most practical way to do processing without huge privacy risks: https://www.anjuna.io/blog/apple-is-using-secure-enclaves-to...
For a short time maybe, but that kind of activity might look like something interesting enough for someone to target you and remove all privacy from your entire existence.
It's a good thing that encrypted data at rest on your local device is inaccessible to cloud based "AI" tools. The problem is that your average person will blithely click "yes/accept/proceed/continue/I consent" on pop up dialogs in a GUI and agree to just about any Terms of Service, including decrypting your data before it's sent to some "cloud" based service.
I see "AI" tools being used even more in the future to permanently tie people to monthly recurring billing services for things like icloud, microsoft's personal grade of office365, google workspace, etc. You'll pay $15 a month forever, and the amount of your data and dependency on the cloud based provider will mean that you have no viable path to ever stop paying it without significant disruption to your life.
I heard that homomorphic encryption can actually preserve all the operations in neural networks, since they are differentiable. Is this true? What is the slowdown in practice?
This is true in principle, yes. In practice, the way this usually works is by converting inputs to bits and bytes, and then computing the result as a digital circuit (AND, OR, XOR).
Doing this encrypted is very slow: without hardware acceleration or special tricks, running the circuit is 1 million times slower than unencrypted, or about 1ms for a single gate. (https://www.jeremykun.com/2024/05/04/fhe-overview/)
When you think about all the individual logic gates involved in just a matrix multiplication, and scale it up to a diffusion model or large transformer, it gets infeasible very quickly.
There are FHE schemes that do better than binary gates (cf. CKKS) but they have other problems in that they require polynomial approximations for all the activation functions. Still they are much better than the binary-FHE schemes for stuff like neural networks, and most hardware accelerators in the pipeline right now are targeting CKKS and similar for this reason.
For some numbers, a ResNet-20 inference can be done in CKKS in like 5 minutes on CPU. With custom changes to the architecture you can get less than one minute, and in my view HW acceleration will improve that by another factor of 10-100 at least, so I'd expect 1s inference of these (still small) networks within the next year or two.
LLMs, however, are still going to be unreasonably slow for a long time.
> Prior to 2011, most cloud-connected devices simply uploaded their data in plaintext.
> Around 2011 our approach to data storage began to evolve. […] began to roll out default end-to-end encryption […] This technology changed the way that keys are managed, to ensure that servers would never see the plaintext content of your messages.
"changed the way that keys are managed" is at a confused contradiction with "uploaded their data in plaintext". If you're going from TLS → E2EE, then yeah, "changed the way keys are managed" miiight make sense, though that's not how I'd phrase it. Then later,
> On the one hand they can (1) send plaintext off to a server, in the process resurrecting many of the earlier vulnerabilities that end-to-end encryption sought to close. Or else (2) they can limit their processing to whatever can be performed on the device itself.
We're still confusing "transmit plaintext" with plaintext being available to the server; the clear option of "use TLS" is omitted. It doesn't really undermine the argument — the server would still have access to the data, and could thus maliciously train AI on it — but it is surprising for a "cryptographer".
> For example, imagine that Apple keeps its promise to deliver messages securely, but then your (Apple) phone goes ahead and uploads (the plaintext) message content to a different set of servers where Apple really can decrypt it. Apple is absolutely using end-to-end encryption in the dullest technical sense… yet is the statement above really accurate? Is Apple keeping its broader promise that it “can’t decrypt the data”?
No, no reasonable person would believe that (though I am sure that if the scenario ever came to be, Apple, or whoever, would likely argue "yes") since it would utterly scuttle the term "E2EE". If you say "Our product supports X", and then have to caveat away 100% of what makes X X, then it's just grift, plain and simple. (Now, whether grift sees regulatory action … well.)
> Now imagine that some other member of the group — not you, but one of your idiot friends — decides to turn on some service that uploads (your) received plaintext messages to WhatsApp.
> In general, what we’re asking here is a question about informed consent.
I would sort of agree, but corporations will expose the consent here to the "friend", and then argue that because the friend consented to your data being uploaded, it is fine. An argument for privacy regulations.
(I don't think you have to go through all this … work. Just upload the user's data. They'll complain, for a bit, but the market has already consolidated into at least an ologopoly, users have shown that, for the most part, they're going to keep using the product rather than leave, or else I'll be ending this comment with a free "2025 will be the Year of the Linux Desktop". What's gonna happen, regulation to ensure a free market remains free¹? Please. Cf. MS Recall, currently in the "complain" phase, but give it time, and we'll reach the "we heard your concerns, and we value your input and take your feedback with the utmost respect ram it down their throats" stage.)
(¹free as in "dictated by the laws of supply & demand", not laissez-faire which is where the US will be headed for the next 4.)
(and … 2011? I'd've said 2013 is when we found out the 4A meant way less than we thought it did, leading to the rise in massive adoption of TLS. Less so E2EE.)
From Apple's document on Advanced Data Protection:
>With Advanced Data Protection enabled, Apple doesn't have the encryption keys needed to help you recover your end-to-end encrypted data.
Apple doesn't have the keys. Somebody else might. Somebody other than you. Also, I think they meant to say decryption keys, although they're probably just dumbing down terminology for the masses.
>If you ever lose access to your account, you’ll need to use one of your account recovery methods
"You'll need to use." Not "there is no way except to use."
>Note: Your account recovery methods are never shared with or known to Apple.
"shared with or known to Apple." Not "shared with or known to anyone else."
The encryption is there, I believe that. I just don't know how many copies of the keys there are. If the only key is with me, it would be super easy for Apple to just say that. I believe that they have said that in the past, but the wording has now changed to this hyper-specific "Apple does not have the key" stuff.
> Although PCC is currently unique to Apple, we can hope that other privacy-focused services will soon crib the idea.
IMHO, Apple's PCC is a step in the right direction in terms of general AI privacy nightmares where they are at today. It's not a perfect system, since it's not fully transparent and auditable, and I do not like their new opt-out photo scanning feature running on PCC, but there really is a lot to be inspired by it.
My startup is going down this path ourselves, building on top of AWS Nitro and Nvidia Confidential Compute to provide end to end encryption from the AI user to the model running on the enclave side of an H100. It's not very widely known that you can do this with H100s but I really want to see this more in the next few years.
Yes, you're correct on both, though I think Google Cloud recently started supporting it as well. AWS will likely have GPU enclave support with Trainium 2 soon (AFAIK, that feature is not publicly offered yet but could be wrong).
We work with Edgeless Systems who manages the GPU enclave on Azure that we speak to from our AWS Nitro instance. While not ideal, the power of enclaves and the attestation verification process, we at least know that we're not leaking privacy by going with a third party GPU enclave provider.
And the most important thing about PCC in my opinion is not the technical aspect (though that's nice) but that Apple views user privacy as something good to be maximized, differing from the view championed by OpenAI and Anthropic (and also adopted by Google and virtually every other major LLM provider by this point) that user interactions must be surveilled for "safety" purposes. The lack of privacy isn't due to a technical limitation--it's intended, and they often brag about it.
Something good to be maximized within the constraints of the systems they have to work within. But at some point with enough compromises it becomes maximizing the perception of privacy, not the reality. Promoting these academic techniques may just be perception management on the part of Apple, if the keys are not controlled solely by the user.
I think this has also a silver lining. The E2E encryption movement especially for messenger apps was largely also used to silently lock users out of their own data and effectively prevent user agency to use their own data to move apps, write automations or archive, this is not just true for whatsapp (the data export feature does not fully work since its launch and was just made to appease some EU law that did not properly check if the button works until the end.) Also signal does not have a way to do this. Maybe with ai coming into the game companies finally decide to provide access to data, I just hope it's in a transparent way with user opt in and user control.
I don't think you can extrapolate a trend from a few apps having bugs in their export code. Google Takeout is also notoriously buggy and they don't use E2E encryption. A more likely explanation is companies of all kinds don't care that much about export functionality, due to the incentives involved.
you CAN extrapolate from nearly all e2e encrypted apps not giving a way to use the data. And there is a big difference between buggy google export features or facebook actively making export unusable to lock in users.
Signal does not have a way to manually export your private keys and chat history, but the process of "moving" your signal account to a new phone is quite straightforward. You put both devices on the same wifi/LAN layer 2 broadcast segment, start the transfer process in the app, input the verification codes displayed on the screen from both devices, and it sends everything over. This moves the private key in a way that does not result in all of your contacts receiving the scary "this person's key has changed" message.
"Moving your account" is not what i talk about, besides not being possible between android and ios with history. User agency means a user is allowed to access his data and do what they want with it how they want, realtime and with whatever code they want to write to do so.
> You might even convince yourself that these questions are “privacy preserving,” since no human police officer would ever rummage through your papers, and law enforcement would only learn the answer if you were (probably) doing something illegal.
Something I've started to see happen but never mentioned is the effect automated detection has on systems: As detection becomes more automated (previously authored algorithms, now with large AI models), there's less cash available for individual case workers, and more trust at the managerial level on automatic detection. This leads to false positives turning into major frustrations since it's hard to get in touch with a person to resolve the issue. When dealing with businesses it's frustrating, but as these get more used in law enforcement, this could be life ruining.
For instance - I got flagged as illegal reviews on Amazon years ago and spent months trying to make my case to a human. Every year or so I try to raise the issue again to leave reviews, but it gets nowhere. Imagine this happening for a serious criminal issue, with the years long back log on some courts, this could ruin someones life.
More automatic detection can work (and honestly, it's inevitable) but it's got to acknowledge that false positives will happen and allocate enough people to resolve those issues. As it stands right now, these detection systems get built and immediately human case workers get laid off, there's this assumption that detection systems REPLACE humans, but it should be that they augment and focus human case workers so you can do more with less - the human aspect needs to be included in the budgeting.
But the incentives aren't there, and the people making the decisions aren't the ones working the actual cases so they aren't confronted with the problem. For them, the question is why save $1m when you could save $2m? With large AI models making it easier and more effective to build automated detection I expect this problem to get significantly worse over the next years.
The UK Post Office scandal is bone-chilling.
Update this to a world where every corner of your life is controlled by a platform monopoly that doesn't even provide the most bare-bones customer service and yeah, this is going to get a lot worse before it gets better.
So if I understand it
1. E2E encryption does work
2. But phones can send plaintext back to the cloud to get help doing AI things
3. And we tend not to know because it’s all “assisstance”
But the solution like anything is pricing. I mean yet again (uber, Airbnb) billions of dollars of VC money is used as subsidy so my photos can get OCR’d.
If phones said “hey for a dollar fifty I can work out what the road sign says behind your dogs head in 32 photos your mum sent you last week” I think we woukd see a different threat landscape
This is - again - unsustainable cash spending distorting markets and common sense. If the market was “we can OCR and analyse these insurance claims” the. Things like privacy and encryption would be first class requirements and harder to sell and build.
By spending a billion they can sell services to people without regulators to ask awkward questions and then they hope step 3. Profit.
I short not even AI can spot patterns in encrypted data, it’s only when plaintext gets sent around in the hope of profit do we see a threat. That’s seems a simple fix if not an easy one
> Yet this approach is obviously much better than what’s being done at companies like OpenAI, where the data is processed by servers that employees (presumably) can log into and access.
No need for presumption here: OpenAI is quite transparent about the fact that they retain data for 30 days and have employees and third-party contractors look at it.
https://platform.openai.com/docs/models/how-we-use-your-data
> To help identify abuse, API data may be retained for up to 30 days, after which it will be deleted (unless otherwise required by law).
https://openai.com/enterprise-privacy/
> Our access to API business data stored on our systems is limited to (1) authorized employees that require access for engineering support, investigating potential platform abuse, and legal compliance and (2) specialized third-party contractors who are bound by confidentiality and security obligations, solely to review for abuse and misuse.
> We are about to face many hard questions about these systems, including some difficult questions about whether they will actually be working for us at all.
And how. I'd lean towards no. Where we're headed feels like XKEYSCORE on steroids. I'd love to take the positive, optimistic bent on this, but when you look at where we've been combined with the behavior of the people in charge of these systems (to be clear, not the researchers or engineers, but c-suite), hope of a neutral, privacy-first future seems limited.
My guess is that the main purpose of agents will be to train the AI on your data. Companies have run out of data on the internet for training AI's, so they'll use agents as an excuse to get access to your personal real-time data. This has always been the business model, you are the product.
Given how politics and companies evolved, I actually trust those people in charge of XKEYSCORE systems more than ever. They may wear suits, but those people usually come from some military background, and have a sense of duty towards defending US, from threats both foreign and domestic, and historically have not really abused their powers no matter what the administration is. XKEYSCORE for example, wasn't really about hacking people, it was just about collecting mass metadata and building profiles, well within the legal system, and the blame should be on the companies that didn't provide privacy tools, because any big government could have build the same system.
Meanwhile, the anti anti-establishment Republican Party since 2016 who cried about big tech turned out to be the biggest pro-establishment fans, giving Elmo an office in a white house and Zucc bending a knee to avoid prosecution.
With these new systems, Id rather have smart people who only work in US defensive forces because of a sense of duty (considering they could get paid much more in the private sector) in charge.
> well within the legal system
It's not a search if we don't find anything, and it's not a seizure if we charge the money with the crime. These are court approved arguments, so they must be correct interpretations.
Point is: modern bureaucrats have proven that they are absolutely willing to abuse power, even in the best of times when there is no real domestic political strife.
The most depressing realization in all of this is that the vast treasure trove of data that we used to have in the cloud thinking it was not scannable even for criminal activity has now become a vector where we shall have thought police coming down upon us for simple ideas of dissent.
A lot of people tried to sound the alarm. It's not "the cloud", it's "other people's computers". And given that other people own these machines, their interests - whether commercial or ideological - will always come first.
To be fair, most people understand that risk. It’s just it is very convenient in a lot of scenarios and some businesses might have not even started without it. Privacy is not that big of a concern for a big chunk of people. And they’re basically voting with their wallets.
> Apple even says it will publish its software images (though unfortunately not the source code) so that security researchers can check them over for bugs.
I think Apple recently changed their stance on this. Now, they say that "source code for certain security-critical PCC components are available under a limited-use license." Of course, would have loved it if the whole thing was open source. ;)
https://github.com/apple/security-pcc/
> The goal of this system is to make it hard for both attackers and Apple employees to exfiltrate data from these devices.
I think Apple is claiming more than that. They are saying 1/ they don't keep any user data (data only gets processed during inference), 2/ no privileged runtime access, so their support engineers can't see user data, and 3/ they make binaries and parts of the source code available to security researchers to validate 1/ and 2/.
You can find Apple PCC's five requirements here: https://security.apple.com/documentation/private-cloud-compu...
Note: Not affiliated with Apple. We read through the PCC security guide to see what an equivalent solution would look like in open source. If anyone is interested in this topic, please hit me up at ozgun @ ubicloud . com.
> Who does your AI agent actually work for?
Yes. I made that point a few weeks ago. The legal concept of principal and agent applies.
Running all content through an AI in the cloud to check for crimethink[1] is becoming a reality. Currently proposed:
- "Child Sexual Abuse Material", which is a growing category that now includes AI-generated images in the US and may soon extend to Japanese animation.
- Threats against important individuals. This may be extended to include what used to be considered political speech in the US.
- Threats against the government. Already illegal in many countries. Bear in mind that Trump likes to accuse people of "treason" for things other than making war against the United States.
- "Grooming" of minors, which is vague enough to cover most interactions.
- Discussing drugs, sex, guns, gay activity, etc. Variously prohibited in some countries.
- Organizing protests or labor unions. Prohibited in China and already searched for.
Note that talking around the issue or jargon won't evade censorship. LLMs can deal with that. Run some ebonics or leetspeak through an LLM and ask it to translate it to standard English. Translation will succeed. The LLM has probably seen more of that dialect than most people.
"If you want a vision of the future, imagine a boot stepping on a face, forever" - Orwell
[1] https://www.orwell.org/dictionary/
A cynic in me is amused at the yet unknown corporation being placed under investigation due to a trigger phrase in one of the meetings transcribed incorrectly.
Your point is worth reiterating.
Or poisoned-data that sets up a trap, so that a system will later confabulate false innocence or guilt when certain topics or targets come up.
The author helpfully emphasized the interesting question at the end
> This future worries me because it doesn’t really matter what technical choices we make around privacy. It does not matter if your model is running locally, or if it uses trusted cloud hardware — once a sufficiently-powerful general-purpose agent has been deployed on your phone, the only question that remains is who is given access to talk to it. Will it be only you? Or will we prioritize the government’s interest in monitoring its citizens over various fuddy-duddy notions of individual privacy.
I do think there are interesting policy questions there. I mean it could hypothetically be mandated that the government must be given access to the agent (in the sense that we and these companies exist in jurisdictions that can pass arbitrary laws; let’s skip the boring and locale specific discussion of whether you think your local government would pass such a law).
But, on a technical level—it seems like it ought to be possible to run an agent locally, on a system with full disk encryption, and not allow anyone who doesn’t have access to the system to talk with it, right? So on a technical level I don’t see how this is any different from where we were previously. I mean you could also run a bunch of regex’s from the 80’s to find whether or not somebody has, whatever, communist pamphlets on their computers.
There’s always been a question of whether the government should be able to demand access to your computer. I guess it is good to keep in mind that if they are demanding access to an AI agent that ran on your computer, they are basically asking for a lossy record of your entire hard drive.
> The author helpfully emphasized the interesting question at the end
We're already there. AI or not doesn't affect the fact that smartphones gather, store, and transmit a great deal of information about their users and their users' actions and interests.
Unreasonable search?
> (in the sense that we and these companies exist in jurisdictions that can pass arbitrary laws; let’s skip the boring and locale specific discussion of whether you think your local government would pass such a law)
Anyway the idea of what’s a reasonable search in the US has been whittled away to almost nothing, right? “The dog smelled weed on your hard drive.” - A cop, probably.
Boring locale specific discussion.
See also CrypTen, Meta's library for privacy preserving machine learning: https://github.com/facebookresearch/CrypTen. This isn't fully homomorphic encryption, but it is multi-party computation (MPC), which hides the inputs from the company owning the model.
But while not revealing user input, it would still reveal the outputs of the model to the company. And yeah, as the article mentions, unfortunately this kind of thing (MPC or fully-homomorphic encryption) probably won't be feasible for the most powerful ML models.
Maybe a little off topic, but is there a way for a distributed app to connect to one of the LLM companies (OpenAI, etc.) without the unencrypted data hitting an in-between proxy server?
An app I'm building uses LLMs to process messages. I don’t want the unencrypted message to hit my server - and ideally I wouldn’t have the ability to decrypt it. But I can’t communicate directly from client -> LLM Service without leaking the API key.
"But I can’t communicate directly from client -> LLM Service without leaking the API key."
There is a way you can do that right now: the OpenAI WebRTC API introduced the idea of an "ephemeral key": https://platform.openai.com/docs/guides/realtime-webrtc
This provides a way for your server to create a limited-time API key for a user which their browser can then use to talk to OpenAI's API directly without proxying through you.
I love this idea, but I want it for way more than just the WebRTC API, and I'd like it for other API providers too.
My ideal version would be a way to create an ephemeral API key that's only allowed to talk to a specific model with a specific pre-baked system prompt (and maybe tool configuration and suchlike) and that only works for a limited time and has a limited token budget.
interesting, will check that out. thanks!
Check out https://www.opaque.co/
Will such processing be cheap enough to be done by a box that plugs into a customers router to handle such? Would they buy them? Notably not just for this use case but others
There is always going to be a gap between local-first processing and what can be achieved in a full-sized datacenter/cloud. That leads to the risks mentioned in the article.
I wrote about Apple's Private Cloud Compute last year; for the foreseeable future, I still think server-side Confidential Computing is the most practical way to do processing without huge privacy risks: https://www.anjuna.io/blog/apple-is-using-secure-enclaves-to...
Is embeddings enough to preserve privacy? If I run the encoder/decoder on device and only communicate with server in embeddings?
No, the original text can largely be recovered from embeddings[0] if you know which embedding model was used.
[0] https://arxiv.org/abs/2310.06816
For a short time maybe, but that kind of activity might look like something interesting enough for someone to target you and remove all privacy from your entire existence.
It's a good thing that encrypted data at rest on your local device is inaccessible to cloud based "AI" tools. The problem is that your average person will blithely click "yes/accept/proceed/continue/I consent" on pop up dialogs in a GUI and agree to just about any Terms of Service, including decrypting your data before it's sent to some "cloud" based service.
I see "AI" tools being used even more in the future to permanently tie people to monthly recurring billing services for things like icloud, microsoft's personal grade of office365, google workspace, etc. You'll pay $15 a month forever, and the amount of your data and dependency on the cloud based provider will mean that you have no viable path to ever stop paying it without significant disruption to your life.
I heard that homomorphic encryption can actually preserve all the operations in neural networks, since they are differentiable. Is this true? What is the slowdown in practice?
This is true in principle, yes. In practice, the way this usually works is by converting inputs to bits and bytes, and then computing the result as a digital circuit (AND, OR, XOR).
Doing this encrypted is very slow: without hardware acceleration or special tricks, running the circuit is 1 million times slower than unencrypted, or about 1ms for a single gate. (https://www.jeremykun.com/2024/05/04/fhe-overview/)
When you think about all the individual logic gates involved in just a matrix multiplication, and scale it up to a diffusion model or large transformer, it gets infeasible very quickly.
There are FHE schemes that do better than binary gates (cf. CKKS) but they have other problems in that they require polynomial approximations for all the activation functions. Still they are much better than the binary-FHE schemes for stuff like neural networks, and most hardware accelerators in the pipeline right now are targeting CKKS and similar for this reason.
For some numbers, a ResNet-20 inference can be done in CKKS in like 5 minutes on CPU. With custom changes to the architecture you can get less than one minute, and in my view HW acceleration will improve that by another factor of 10-100 at least, so I'd expect 1s inference of these (still small) networks within the next year or two.
LLMs, however, are still going to be unreasonably slow for a long time.
TFA makes some rather basic errors.
First,
> Prior to 2011, most cloud-connected devices simply uploaded their data in plaintext.
> Around 2011 our approach to data storage began to evolve. […] began to roll out default end-to-end encryption […] This technology changed the way that keys are managed, to ensure that servers would never see the plaintext content of your messages.
"changed the way that keys are managed" is at a confused contradiction with "uploaded their data in plaintext". If you're going from TLS → E2EE, then yeah, "changed the way keys are managed" miiight make sense, though that's not how I'd phrase it. Then later,
> On the one hand they can (1) send plaintext off to a server, in the process resurrecting many of the earlier vulnerabilities that end-to-end encryption sought to close. Or else (2) they can limit their processing to whatever can be performed on the device itself.
We're still confusing "transmit plaintext" with plaintext being available to the server; the clear option of "use TLS" is omitted. It doesn't really undermine the argument — the server would still have access to the data, and could thus maliciously train AI on it — but it is surprising for a "cryptographer".
> For example, imagine that Apple keeps its promise to deliver messages securely, but then your (Apple) phone goes ahead and uploads (the plaintext) message content to a different set of servers where Apple really can decrypt it. Apple is absolutely using end-to-end encryption in the dullest technical sense… yet is the statement above really accurate? Is Apple keeping its broader promise that it “can’t decrypt the data”?
No, no reasonable person would believe that (though I am sure that if the scenario ever came to be, Apple, or whoever, would likely argue "yes") since it would utterly scuttle the term "E2EE". If you say "Our product supports X", and then have to caveat away 100% of what makes X X, then it's just grift, plain and simple. (Now, whether grift sees regulatory action … well.)
> Now imagine that some other member of the group — not you, but one of your idiot friends — decides to turn on some service that uploads (your) received plaintext messages to WhatsApp.
> In general, what we’re asking here is a question about informed consent.
I would sort of agree, but corporations will expose the consent here to the "friend", and then argue that because the friend consented to your data being uploaded, it is fine. An argument for privacy regulations.
(I don't think you have to go through all this … work. Just upload the user's data. They'll complain, for a bit, but the market has already consolidated into at least an ologopoly, users have shown that, for the most part, they're going to keep using the product rather than leave, or else I'll be ending this comment with a free "2025 will be the Year of the Linux Desktop". What's gonna happen, regulation to ensure a free market remains free¹? Please. Cf. MS Recall, currently in the "complain" phase, but give it time, and we'll reach the "we heard your concerns, and we value your input and take your feedback with the utmost respect ram it down their throats" stage.)
(¹free as in "dictated by the laws of supply & demand", not laissez-faire which is where the US will be headed for the next 4.)
(and … 2011? I'd've said 2013 is when we found out the 4A meant way less than we thought it did, leading to the rise in massive adoption of TLS. Less so E2EE.)
From Apple's document on Advanced Data Protection:
>With Advanced Data Protection enabled, Apple doesn't have the encryption keys needed to help you recover your end-to-end encrypted data.
Apple doesn't have the keys. Somebody else might. Somebody other than you. Also, I think they meant to say decryption keys, although they're probably just dumbing down terminology for the masses.
>If you ever lose access to your account, you’ll need to use one of your account recovery methods
"You'll need to use." Not "there is no way except to use."
>Note: Your account recovery methods are never shared with or known to Apple.
"shared with or known to Apple." Not "shared with or known to anyone else."
The encryption is there, I believe that. I just don't know how many copies of the keys there are. If the only key is with me, it would be super easy for Apple to just say that. I believe that they have said that in the past, but the wording has now changed to this hyper-specific "Apple does not have the key" stuff.
> Although PCC is currently unique to Apple, we can hope that other privacy-focused services will soon crib the idea.
IMHO, Apple's PCC is a step in the right direction in terms of general AI privacy nightmares where they are at today. It's not a perfect system, since it's not fully transparent and auditable, and I do not like their new opt-out photo scanning feature running on PCC, but there really is a lot to be inspired by it.
My startup is going down this path ourselves, building on top of AWS Nitro and Nvidia Confidential Compute to provide end to end encryption from the AI user to the model running on the enclave side of an H100. It's not very widely known that you can do this with H100s but I really want to see this more in the next few years.
I didn't actually realize that AWS supported this, I thought Azure was the only one offering it (https://azure.microsoft.com/en-us/blog/azure-confidential-co...)
Are you speaking of this functionality? https://developer.nvidia.com/blog/confidential-computing-on-... (and am I just failing to find the relevant AWS docs?)
Yes, you're correct on both, though I think Google Cloud recently started supporting it as well. AWS will likely have GPU enclave support with Trainium 2 soon (AFAIK, that feature is not publicly offered yet but could be wrong).
We work with Edgeless Systems who manages the GPU enclave on Azure that we speak to from our AWS Nitro instance. While not ideal, the power of enclaves and the attestation verification process, we at least know that we're not leaking privacy by going with a third party GPU enclave provider.
And the most important thing about PCC in my opinion is not the technical aspect (though that's nice) but that Apple views user privacy as something good to be maximized, differing from the view championed by OpenAI and Anthropic (and also adopted by Google and virtually every other major LLM provider by this point) that user interactions must be surveilled for "safety" purposes. The lack of privacy isn't due to a technical limitation--it's intended, and they often brag about it.
Something good to be maximized within the constraints of the systems they have to work within. But at some point with enough compromises it becomes maximizing the perception of privacy, not the reality. Promoting these academic techniques may just be perception management on the part of Apple, if the keys are not controlled solely by the user.
"The goal of encryption is to ensure that only two parties, the receiver and sender, are aware of the contents of your data.
Thus, AI training on your data breaks this, because it's another party.
You now don't have encryption."
Thanks for coming to my blah blah blah
The article has nothing to do with model training.
I think this has also a silver lining. The E2E encryption movement especially for messenger apps was largely also used to silently lock users out of their own data and effectively prevent user agency to use their own data to move apps, write automations or archive, this is not just true for whatsapp (the data export feature does not fully work since its launch and was just made to appease some EU law that did not properly check if the button works until the end.) Also signal does not have a way to do this. Maybe with ai coming into the game companies finally decide to provide access to data, I just hope it's in a transparent way with user opt in and user control.
1. Is data encrypted in transit?
2. Can the user access their data at rest?
Those two things are entirely orthogonal.
I don't think you can extrapolate a trend from a few apps having bugs in their export code. Google Takeout is also notoriously buggy and they don't use E2E encryption. A more likely explanation is companies of all kinds don't care that much about export functionality, due to the incentives involved.
you CAN extrapolate from nearly all e2e encrypted apps not giving a way to use the data. And there is a big difference between buggy google export features or facebook actively making export unusable to lock in users.
Signal does not have a way to manually export your private keys and chat history, but the process of "moving" your signal account to a new phone is quite straightforward. You put both devices on the same wifi/LAN layer 2 broadcast segment, start the transfer process in the app, input the verification codes displayed on the screen from both devices, and it sends everything over. This moves the private key in a way that does not result in all of your contacts receiving the scary "this person's key has changed" message.
"Moving your account" is not what i talk about, besides not being possible between android and ios with history. User agency means a user is allowed to access his data and do what they want with it how they want, realtime and with whatever code they want to write to do so.