Reading the whitepaper, the inference provider still has the ability to access the prompt and response plaintext. This scheme does seem to guarantee that plaintext cannot be read for all other parties (e.g. the API router), and that the client's identity is hidden and cannot be associated with their request. Perhaps the precise privacy guarantees and allowances should be summarized in the readme.
With that in mind, does this scheme offer any advantage over the much simpler setup of a user sending an inference request:
- directly to an inference provider (no API router middleman)
- that accepts anonymous crypto payments (I believe such things exist)
- using a VPN to mask their IP?
Howdy, head of Eng at confident.security here, so excited to see this out there.
I'm not sure I understand what you mean by inference provider here? The inference workload is not shipped off the compute node once it's been decrypted to e.g. OpenAI, it's running directly on the compute machine on open source models loaded there. Those machines are cryptographically attesting to the software they are running. Proving, ultimately, that there is no software that is logging sensitive info off the machine, and the machine is locked down, no SSH access.
This is how Apple's PCC does it as well, clients of the system will not even send requests to compute nodes that aren't making these promises, and you can audit the code running on those compute machines to check that they aren't doing anything nefarious.
The privacy guarantee we are making here is that no one, not even people operating the inference hardware, can see your prompts.
Thanks for the reply! By "inference provider" I meant someone operating a ComputeNode. I initially skimmed the paper, but I've now read more closely and see that we're trying to get guarantees that even a malicious operator is unable to e.g. exfiltrate prompt plaintext.
Despite recent news of vulnerabilities, I do think that hardware-root-of-trust will eventually be a great tool for verifiable security.
A couple follow-up questions:
1. For the ComputeNode to be verifiable by the client, does this require that the operator makes all source code running on the machine publicly available?
2. After a client validates a ComputeNode's attestation bundle and sends an encrypted prompt, is the client guaranteed that only the ComputeNode running in its attested state can decrypt the prompt? Section 2.5.5 of the whitepaper mentions expiring old attestation bundles, so I wonder if this is to protect against a malicious operator presenting an attestation bundle that doesn't match what's actually running on the ComputeNode.
> no one, not even people operating the inference hardware
You need to be careful with these claims IMO. I am not involved directly in CoCo so my understanding lacks nuance but after https://tee.fail I came to understand that basically there's no HW that actually considers physical attacks in scope for their threat model?
The Ars Technica coverage of that publication has some pretty yikes contrasts between quotes from people making claims like yours, and the actual reality of the hardware features.
https://arstechnica.com/security/2025/10/new-physical-attack...
My current understanding of the guarantees here is:
- even if you completely pwn the inference operator, steal all root keys etc, you can't steal their customers' data as a remote attacker
- as a small cabal of arbitrarily privileged employees of the operator, you can't steal the customers' data without a very high risk of getting caught
- BUT, if the operator systematically conspires to steal the customers' data, they can. If the state wants the data and is willing to spend money on getting it, it's theirs.
I'm happy to be careful, you are right we are relying on TEEs and vTPMs as roots of trust here and TEEs have been compromised by attackers with physical access.
This is actually part of why we think it's so important to have the non-targetability part of the security stack as well, so that even if someone where to physically compromise some machines at a cloud provider, there would be no way for them to reliably route a target's requests to that machine.
> The privacy guarantee we are making here is that no one, not even people operating the inference hardware, can see your prompts.
that cannot be met, period. your asssumptions around physical protections are invalid or at least incorrect. It works for Apple (well enough) because of the high trust we place in their own physical controls, and market incentive to protect that at all costs.
> This is how Apple's PCC does it as well [...] and you can audit the code running on those compute machines to check that they aren't doing anything nefarious.
just based on my recollection, and I'm not going to have a new look at it to validate what I'm saying here, but with PCC, no you can't actually do that. With PCC you do get an attestation, but there isn't actually a "confidential compute" aspect where that attestation (that you can trust) proves that is what is running. You have to trust Apple at that lowest layer of the "attestation trust chain".
I feel like with your bold misunderstandings you are really believing your own hype. Apple can do that, sure, but a new challenger cannot. And I mean your web page doesn't even have an "about us" section.
That's a strong claim for not looking into it at all.
From a brief glance at the white paper it looks like they are using TEE, which would mean that the root of trust is the hardware chip vendor (e.g. Intel). Then, it is possible for confidentiality guarantees to work if you can trust the vendor of the software that is running. That's the whole purpose of TEE.
It doesn’t matter either way. Intel is an American company as well, and thus unsuitable as a trust root.
A company of what country would you prefer?
Everyone likes to dunk on the US, but I doubt you could provide a single example of a country that is certainly a better alternative (to be clear I believe many of the west up in the same boat).
> the inference provider still has the ability to access the prompt and response plaintext
Folks may underestimate the difficulty of providing compute that the provider “cannot”* access to reveal even at gunpoint.
BYOK does cover most of it, but oh look, you brought me and my code your key, thanks… Apple's approach, and certain other systems such as AWS's Nitro Enclaves, aim at this last step of the problem:
- https://security.apple.com/documentation/private-cloud-compu...
- https://aws.amazon.com/confidential-computing/
NCC Group verified AWS's approach and found:
1. There is no mechanism for a cloud service provider employee to log in to the underlying host.
2. No administrative API can access customer content on the underlying host.
3. There is no mechanism for a cloud service provider employee to access customer content stored on instance storage and encrypted EBS volumes.
4. There is no mechanism for a cloud service provider employee to access encrypted data transmitted over the network.
5. Access to administrative APIs always requires authentication and authorization.
6. Access to administrative APIs is always logged.
7. Hosts can only run tested and signed software that is deployed by an authenticated and authorized deployment service. No cloud service provider employee can deploy code directly onto hosts.
- https://aws.amazon.com/blogs/compute/aws-nitro-system-gets-i...
Points 1 and 2 are more unusual than 3 - 7.
Folks who enjoy taking things apart to understand them can hack at Apple's here:
https://security.apple.com/blog/pcc-security-research/
* Except by, say, withdrawing the system (see Apple in UK) so users have to use something less secure, observably changing the system, or other transparency trippers.
Yes but at the end of the day you need to trust the cloud provider tools which expands the trust boundary from just hardware root of trust. Who is to guarantee they will not create a malicious tool update and push it then retract it? It is nowhere captured and you cannot prove it.
You can detect and prove it because the hardware attestation signature will change.
You might not know what change was made, or have any prior warning of the change. But you will be able to detect it happening. Which means an operator only gets to play that card once, after which nobody will trust them again.
My logic is that these "confidential compute" problems suffer from some of the same issues as "immutable storage in blockchain".
I.e.: If the security/privacy guarantees really are as advertised, then ipso facto someone could store child porn in the system and the provider couldn't detect this.
Then by extension, any truly private system is exposing themselves to significant business, legal, and moral risk of being tarred and feathered along with the pedos that used their system.
It's a real issue, and has come up regularly with blockchain based data storage. If you make it "cencorship proof", the by definition you can't scrub it of illegal data!
Similarly, if cloud providers allow truly private data hosting, then they're exposing themselves to the risk of hosting data that is being stored with that level of privacy guarantees precisely because it is so very, very illegal.
(Or substitute: Stolen state secrets that will have the government come down on you like a ton of bricks. Stolen intellectual property. Blackmail information on humourless billionaires. Illegal gambling sites. Nuclear weapons designs. So on, and so forth.)
This is hardly a new problem that only appears in the cloud. Any bank that offers a private secure storage facility I.e. a safety deposit box, or anyone that offers a PO Box service is also exposed to the same risk.
But both of these services exist, and have existed for hundreds of years, and don’t require service providers to go snooping though their customer’s possessions or communications.
> I.e.: If the security/privacy guarantees really are as advertised, then ipso facto someone could store child porn in the system and the provider couldn't detect this.
But what they would be storing in this case is not illegal content. Straight up. Encrypted bits without a key are meaningless.
There is nothing stopping a criminal from uploading illegal content to Google drive as an encrypted blob. There's nothing Google can do about it, and there is no legal repercussion (to my knowledge) of holding such a blob.
At the end if the day, Nitro Enclaves are still “trust Amazon”, which is a poor guarantee. NVIDIA+AMD offers hardware backed enclave features for their GPUs which is the superior solution here.
Aren’t they both hardware backed, just changing the X in “trust X”?
be sure to let us know when you can run eg nginx on a GPU in said enclave.
> Folks may underestimate the difficulty of providing compute that the provider “cannot”* access to reveal even at gunpoint.
It's even harder to do this plus the hard requirement of giving the NSA access.
Or alternatively, give the user a verifiable guarantee that nobody has access.
at that point, it seems easier to run a slightly worse model locally. (or on a rented server)
Which is apples own approach until the compute requirements need them to run some compute on cloud.
Just a shame they spent so long skimping on iPhone memory. The tail-end of support for 4gb and 6gb handsets is going to push that compute barrier pretty low.
It's probably illegal for a business to take anonymous cryptocurrency payments in the EU. Businesses are allowed to take traceable payments only, or else it's money laundering.
With the caveat that it's not clear what precisely is illegal about these payments and to what level it's illegal. It might be that a business isn't allowed to have any at all, or isn't allowed to use them for business, or can use them for business but can't exchange them for normal currency, or can do all that but has to check their customer's passport and fill out reams of paperwork.
https://bitcoinblog.de/2025/05/05/eu-to-ban-trading-of-priva...
We are introducing Verifiably Private AI [1] which actually solves all of the issues you mention. Everything across the entire chain is verifiably private (or in other words, transparent to the user in such a way they can verify what is running across the entire architecture).
It should be able to support connecting via an OpenPCC client, then!
Whitepaper?
I was part of a team that does the same thing. Arguably as a paid service, but source availability and meaningful attestation.
Service: https://www.privatemode.ai/ Code: https://github.com/edgelesssys/privatemode-public
OpenPCC is Apache 2.0 without a CLA to prevent rugpulls whereas edgeless is BSL
<3
Exactly, attestation is what matters. Excluding the inference provider from the prompt is the USP here. Privatemode can do that via an attestation chain (source code -> reproducible build -> TEE attestation report) + code/stack that ensures isolation (Kata/CoCo, runtime policy).
Yes: "provably" private... unless you have $1000 for a logic analyzer and a steady hand to solder together a fake DDR module.
well, also indefinite time and physical access.
Which is what the provider themselves have, by definition. The people who run these services are literally sitting next to the box day in and day out... this isn't "provably" anything. You can trust them not to take advantage of the fact that they own the hardware, and you can even claim it makes it ever so slightly harder for them to do so, but this isn't something where the word "provably" is anything other than a lie.
yeah, for a moment I was reading it as being a holomorphic encryption type setup, which I think is the only case where you can say 'provably private'.
It's better than nothing, I guess...
But if you placed the server at the NSA, and said "there is something on here that you really want, it's currently powered on and connected to the network, and the user is accessing it via ssh", it seems relatively straightforward for them to intercept and access.
If you trust the provider then it does not make it much better to use such architecture. If you do not then at least the execution should be inside a confidential system so that even soldering would not get you to data
impressive work jmo - thanks for open sourcing this (and OSI-compliant)
we are working on a challenge which is somewhat like a homomorphic encryption problem - I'm wondering if OpenPCC could help in some way? :
When developing websites/apps, developers generally use logs to debug production issues. However with wearables, logs can be privacy issue: imagine some AR glasses logging visual data (like someone's face). Would OpenPCC help to extract/clean/anonymize this sort of data for developers to help with their debugging?
Yep, you could run an anonymization workload inside the OpenPCC compute node. We target inference as the "workload" but it's really just attested HTTP server where you can't see inside. So, in this case your client (the wearable) would send its data first through OpenPCC to a server that runs some anonymization process.
If it's possible to anonymize on the wearable, that would be simpler.
The challenge is what does the anonymizer "do" to be perfect?
As an aside, IMO homomorphic encryption (still) isn't ready...
Really nice release. Excited to see this out in the wild and hopeful more companies leverage this for better end user privacy.
Quite similar to what Azure with conf ai inference did [1].
[1] https://techcommunity.microsoft.com/blog/azureconfidentialco...
I haven’t been able to find their source code. Pretty important for the transparency side of it. Have you seen it?
Glad to see Golang here. Go will surpass Python in the AI field, mark my words.
That's nice... in theory. Like it could be cool, and useful... but like what would I actually run on it if I'm not a spammer?
Edit : reminds me of federated learning and FlowerLLM (training only AFAIR, not inference), like... yes, nice, I ALWAYS applaud any way to disentangle from proprieaty software and wall gardens... but like what for? What actual usage?
Gimme an actual example instead of downvoting, help me learn.
Edit on that too : makes me think of OpenAI Whisper as a service via /e/OS and supposedly anonymous proxying (by mixing), namely running STT remotely. That would be an actual potential usage... but IMHO that's low end enough to be run locally. So I'm still looking for an application here.
Are you looking for a general application of LLMs too large to run locally? Because anything you might use remote inference for, you might want to use privately.
> would I actually run on it if I'm not a spammer?
> Gimme an actual example instead of downvoting, help me learn.
Basically you asked a bunch of people on a privacy minded forum, why should they be allowed to encrypt their data? What are you (they) hiding!? Are you a spammer???
Apple is beloved for their stance on privacy, and you basically called everyone who thinks that's more than marketing, a spammer. And before you start arguing no you didn't, it doesn't matter that you didn't, what matters is that that's how your comment made people feel. You can say they're the stupid ones because that's not what you wrote, but if you're genuinely asking for feedback about the downvotes, there you are.
You seriously can't imagine any reason to want to use an LLM privately other than to use it to write spam bots and to spam people? At the very least expand your scope past spamming to, like, also using it to write ransomware.
The proprietary models that can't be run locally are SOTA and local models, even if they can come close, simply aren't what people want.
Where is the compute node source code?
Thought this was going to be about Orchard from the title.
@dang can we modify the title to acknowledge that it's specific to chatbots? The title reads like this is about generic compute, and the content is emphatically not about generic compute.
I realize this is just bad branding by apple but it's still hella confusing.
It does work generically. Like Apple, we initially targeted inference, but it under the hood just an anonymous, attested HTTP server wrapper. The ComputeNode can run an arbitrary workload.
Interesting!