Ok, I am sorry for the ad hominem shenanigans from my side. I was frustrated because your post wasn’t specific, not even with one example, and without specifics (a quote, a ref) I end up piling up benefit of the doubt and we talk past each other. So here are mine.
No, I did not mean Dual EC. You’re right that it’s a completely different thing in form and execution. I meant Snowden’s own definition, from his tweets of 28 Oct 2015. He was answering UK minister Joanna Shields, who had said it’s “not about creating back doors” but about companies giving access “when presented with a warrant”. Snowden replied:
“A backdoor is a mechanism to provide secret access to otherwise confidential communications. It has nothing to do with warrants.”
“The warrant is immaterial. A backdoor that requires a warrant is still a backdoor.”
That’s more about secret access to confidential stuff than about weakening crypto. So my point was never “the graph = Dual_EC”. It was: how close does the graph come to Snowden’s interpretation? Frankly: not all the way. Nobody has shown a third party actually querying it, so by his definition it’s not a backdoor. At most it’s the room behind the door, already prepared and even made resilient.
Regarding these points I agree with you:
Sleep/reachability/etc. tagging from timestamps is trivial. Your journalctl example is perfect.
A lot of it is the advertised product (search “Duane”, Focus/Sleep).
The “applies to any personalized assistant” line is buried. Though I’d read it as a bigger problem, not a smaller one: it’s a class issue, not an Apple issue. It’s also maybe, at least my fear, where the privacy increasing momentum peaking with GDPR is being asphyxiated. People become even more transparent using AI and that data became even more valuable. Now maybe my tinfoil hat take: Look at India using their insane numbers as an experimental platform for rent to several big techs in a (parallel/‘kind of similar’ methods) competition i. e. with CCP. The problem is not enough data for training, right, so the forces against privacy increase tenfold, similar to AI intelligence increase accelerating in a way, the labs actually hide as much as possible (see Anthropic IPO redactions). Tinfoil hat off.
Where I don’t:
Outside encryption. Messages are E2EE in transit, but the graph is the already analysed result, sitting in plaintext SQLite (“stored in plaintext but located in a zone protected by TCC Full Disk Access”, states the blog article). Whoever gets into the unlocked device (malware, physical access, forensic tools, the article names iLEAPP) gets a ready-made profile instead of raw data.
Your logs comparison actually shows the difference. On my Linux I can read journalctl, rotate it, vacuum it. Here the article says Apple provides “no general way to correct an inference that is wrong”, and when you delete a contact the row moves to a second table, expired_stable_graph, “and keeps its confidence and its sources. Like wtf: “Dear customer, you want to delete some meta data? No problem, we make a copy first and then delete it, but lets forget about the copy. Have a beautiful day, nothing to see here…”. How long it stays there is not clear”. So even if I, hypothetically the owner, wanted to respect someone else’s privacy, I can’t really and but even if I thought I did I get gaslighted. That’s hopefully AND hopefully not “the OS working as advertised”. Maybe it depends on who one is leaning team FBI’ish or more towards grandma~average citizen.
Non-users. People who never bought Apple (like me alledgetly) get relationship labels, contact patterns and face clusters on every Apple device of everyone they ‘talk’. This is so infuriating. Your dmesg doesn’t do that by default to your contacts.
And to the thread question (is there evidence of backdoors on iOS): the closest documented cases are Zdziarski 2014 (undocumented services like file_relay that bypassed backup encryption, Apple admitted they exist but called them diagnostics) and Operation Triangulation 2023 (undocumented hardware feature used to bypass kernel memory protection, CVE-2023-38606, intent unknown). Neither proves a deliberate door for a third party. Same for the graph: undocumented and intentional, yes. Access by someone other than the owner, not shown.
So maybe we agree it’s not clickbait-level “Apple backdoor!”, but not just downplayed “wtf logs exist? /s” either.
No worries, I think you’re making a lot of really good points that need to be examined!
I think the graph is even farther away from the Snowden backdoor definition tweet than you say. It seems to be a room with only a front door, protected by all sorts of Byzantine security measures but not obfuscated from the authorized user (in plaintext, which I don’t 100% agree with but there are actually some good reasons to not encrypt it. In the Jersey mafioso Hitman gameplay commentary “yeah, dats not what I wooda done but you can’t argue wit results”).
I think history will look back on gdpr in the same way as it will (lowkey presently does) view the American ppaca: a wavebreak to disrupt people’s expression of their needs, funnel money to companies implementing it and send information to intelligence services and police.
As far as India goes I really try to step lightly there but that’s not really tinfoil hat territory. It never developed past being a colony, just transitioned to the subject of neocolonialism, and what is done there eventually makes its way to the anglosphere.
What may not be made clear in the article is that full disk access is gated by a (if I remember right) two step process that the authenticated user has to go through, so it’s not even just behind a uac popup, but a whole rigmarole the user has to put their password in to even turn on.
It’s one of the reasons to never give your phone or computer to cops and if they take it never to give them the passwords or passkeys and keep biometric locks disabled. Access to your social media through the logged in device user account gives the same information as the graph in the article.
Users not being able to take fine grained control over what metadata gets associated with them is a big deal (and a really good example of a difference between systemd logs and the graph!) but I also don’t know exactly how that would work if they could. Expectations seem to draw a picture of a radio button or slider, not a database viewing tool. Idk.
The delinking as opposed to deleting doesn’t alarm me at all. That’s how filesystems work, it’s how the systems operation is presented to the user in the form of the trash can or recycle bin metaphor, it’s no surprise that it gets handled the same way in a database. Heck, there was some best practices manual for sql use I read at least twenty years ago that had delinking the row as best practices for deletion in production environments. It just seems more like normal operating procedure as opposed to malice.
As far as people who never bought into the apple ecosystem being linked to a macos users graph: the same has been true of vCard contacts, email, all social media but specifically Facebook tagging which builds associations internally for matching tags of even unregistered nonusers, push notifications and a bunch of other stuff I’m not even thinking about. It’s one of those cows out of the barn situations except they got out around forty years ago…
I’m possibly more tinfoil hatted on operation triangulation than anyone else. You wanna exchange paranoid delusions about that you’re talking to the right guy.
Arm CPUs arent clean room designs, they’re a collection of silicon bits and bobs that are licensed from companies that design them and sell them to fabless companies. As such there are two vectors that mmio can get in: a designer leaving some way to rescue a chip that the hapless integrator didn’t properly hook up or a designer purposefully setting a backdoor with malicious intent.
Designers of licensable arm functions are in the same position as newmans character in the movie Jurassic park. Nobody understands what they’re doing and and if they have the guile to get an exploit past the person who buys their design and the disgruntlement to use it (or the desire for money to sell it to someone who will) then there’s nothing stopping them.
The licensed function ic world was never under serious scrutiny until what I’d call the middle of the smartphone lifecycle, now, and some of the industry fallout from ot has been really showing that. See apples move to emte and others dragging their feet.
I think it’s a hard sell to call ot an apple backdoor though, just like you, because of the crazy chain required to even be able to make a mmio request.
I guess I still hate that the article makes such exaggerated claims that appeal to seratonin maxing paranoiacs when sober minded consideration of risk would get more people to make better choices on a platform like macos/ios that I feel offers pretty good controls and processes to enable those choices.
Ok, I am sorry for the ad hominem shenanigans from my side. I was frustrated because your post wasn’t specific, not even with one example, and without specifics (a quote, a ref) I end up piling up benefit of the doubt and we talk past each other. So here are mine.
No, I did not mean Dual EC. You’re right that it’s a completely different thing in form and execution. I meant Snowden’s own definition, from his tweets of 28 Oct 2015. He was answering UK minister Joanna Shields, who had said it’s “not about creating back doors” but about companies giving access “when presented with a warrant”. Snowden replied:
“A backdoor is a mechanism to provide secret access to otherwise confidential communications. It has nothing to do with warrants.” “The warrant is immaterial. A backdoor that requires a warrant is still a backdoor.”
a source URL : https://www.kitguru.net/gaming/security-software/jon-martindale/snowden-blasts-british-government-over-contradiction-on-encryption/
That’s more about secret access to confidential stuff than about weakening crypto. So my point was never “the graph = Dual_EC”. It was: how close does the graph come to Snowden’s interpretation? Frankly: not all the way. Nobody has shown a third party actually querying it, so by his definition it’s not a backdoor. At most it’s the room behind the door, already prepared and even made resilient.
Regarding these points I agree with you:
Where I don’t:
Outside encryption. Messages are E2EE in transit, but the graph is the already analysed result, sitting in plaintext SQLite (“stored in plaintext but located in a zone protected by TCC Full Disk Access”, states the blog article). Whoever gets into the unlocked device (malware, physical access, forensic tools, the article names iLEAPP) gets a ready-made profile instead of raw data.
Your logs comparison actually shows the difference. On my Linux I can read journalctl, rotate it, vacuum it. Here the article says Apple provides “no general way to correct an inference that is wrong”, and when you delete a contact the row moves to a second table, expired_stable_graph, “and keeps its confidence and its sources. Like wtf: “Dear customer, you want to delete some meta data? No problem, we make a copy first and then delete it, but lets forget about the copy. Have a beautiful day, nothing to see here…”. How long it stays there is not clear”. So even if I, hypothetically the owner, wanted to respect someone else’s privacy, I can’t really and but even if I thought I did I get gaslighted. That’s hopefully AND hopefully not “the OS working as advertised”. Maybe it depends on who one is leaning team FBI’ish or more towards grandma~average citizen.
Non-users. People who never bought Apple (like me alledgetly) get relationship labels, contact patterns and face clusters on every Apple device of everyone they ‘talk’. This is so infuriating. Your dmesg doesn’t do that by default to your contacts.
And to the thread question (is there evidence of backdoors on iOS): the closest documented cases are Zdziarski 2014 (undocumented services like file_relay that bypassed backup encryption, Apple admitted they exist but called them diagnostics) and Operation Triangulation 2023 (undocumented hardware feature used to bypass kernel memory protection, CVE-2023-38606, intent unknown). Neither proves a deliberate door for a third party. Same for the graph: undocumented and intentional, yes. Access by someone other than the owner, not shown.
So maybe we agree it’s not clickbait-level “Apple backdoor!”, but not just downplayed “wtf logs exist? /s” either.
No worries, I think you’re making a lot of really good points that need to be examined!
I think the graph is even farther away from the Snowden backdoor definition tweet than you say. It seems to be a room with only a front door, protected by all sorts of Byzantine security measures but not obfuscated from the authorized user (in plaintext, which I don’t 100% agree with but there are actually some good reasons to not encrypt it. In the Jersey mafioso Hitman gameplay commentary “yeah, dats not what I wooda done but you can’t argue wit results”).
I think history will look back on gdpr in the same way as it will (lowkey presently does) view the American ppaca: a wavebreak to disrupt people’s expression of their needs, funnel money to companies implementing it and send information to intelligence services and police.
As far as India goes I really try to step lightly there but that’s not really tinfoil hat territory. It never developed past being a colony, just transitioned to the subject of neocolonialism, and what is done there eventually makes its way to the anglosphere.
What may not be made clear in the article is that full disk access is gated by a (if I remember right) two step process that the authenticated user has to go through, so it’s not even just behind a uac popup, but a whole rigmarole the user has to put their password in to even turn on.
It’s one of the reasons to never give your phone or computer to cops and if they take it never to give them the passwords or passkeys and keep biometric locks disabled. Access to your social media through the logged in device user account gives the same information as the graph in the article.
Users not being able to take fine grained control over what metadata gets associated with them is a big deal (and a really good example of a difference between systemd logs and the graph!) but I also don’t know exactly how that would work if they could. Expectations seem to draw a picture of a radio button or slider, not a database viewing tool. Idk.
The delinking as opposed to deleting doesn’t alarm me at all. That’s how filesystems work, it’s how the systems operation is presented to the user in the form of the trash can or recycle bin metaphor, it’s no surprise that it gets handled the same way in a database. Heck, there was some best practices manual for sql use I read at least twenty years ago that had delinking the row as best practices for deletion in production environments. It just seems more like normal operating procedure as opposed to malice.
As far as people who never bought into the apple ecosystem being linked to a macos users graph: the same has been true of vCard contacts, email, all social media but specifically Facebook tagging which builds associations internally for matching tags of even unregistered nonusers, push notifications and a bunch of other stuff I’m not even thinking about. It’s one of those cows out of the barn situations except they got out around forty years ago…
I’m possibly more tinfoil hatted on operation triangulation than anyone else. You wanna exchange paranoid delusions about that you’re talking to the right guy.
Arm CPUs arent clean room designs, they’re a collection of silicon bits and bobs that are licensed from companies that design them and sell them to fabless companies. As such there are two vectors that mmio can get in: a designer leaving some way to rescue a chip that the hapless integrator didn’t properly hook up or a designer purposefully setting a backdoor with malicious intent.
Designers of licensable arm functions are in the same position as newmans character in the movie Jurassic park. Nobody understands what they’re doing and and if they have the guile to get an exploit past the person who buys their design and the disgruntlement to use it (or the desire for money to sell it to someone who will) then there’s nothing stopping them.
The licensed function ic world was never under serious scrutiny until what I’d call the middle of the smartphone lifecycle, now, and some of the industry fallout from ot has been really showing that. See apples move to emte and others dragging their feet.
I think it’s a hard sell to call ot an apple backdoor though, just like you, because of the crazy chain required to even be able to make a mmio request.
I guess I still hate that the article makes such exaggerated claims that appeal to seratonin maxing paranoiacs when sober minded consideration of risk would get more people to make better choices on a platform like macos/ios that I feel offers pretty good controls and processes to enable those choices.