Hacker Newsnew | past | comments | ask | show | jobs | submit | comex's commentslogin

I have Claude and ChatGPT's desktop apps installed and neither has asked me for Full Disk Access. It's Meta whose AI app was, at least allegedly, directing users to enable Full Disk Access in order to access the Messages app.

To be fair, there is/was no way to request access to Messages specifically. And Claude and ChatGPT do both have Messages connectors, but you have to go out of your way to search for them. Presumably the AI itself will also tell you about Full Disk Access if you ask for something that requires it.

(Also, both of those apps do promote 'computer use', i.e. letting the AI remote-desktop your computer, which effectively grants full system control anyway. But at least in that case it's more obvious to the user what's going on.)


> It's Meta whose AI app was, at least allegedly, directing users to enable Full Disk Access in order to access the Messages app.

What the actual fuck? A messaging app has no business having full-disk access. I can only assume the real reason for this is to manipulate the user into letting them read files they have no business reading. I would consider that spyware.

How is this not getting them banned from AppStore?


You’re misunderstanding. The Messages (Apple’s app) database is only accessible to apps that have full disk access. If you’d like your agent to operate on those then you need to grant it this permission. Muse is also not distributed on the App Store on macOS.

You misunderstood the parent post. Prompting for the documents folder is one way that "[y]ou can access files without full disk access". With full disk access, you get almost everything.

Though not quite everything. For truly full disk access you need to disable SIP.


Yes, sorry. I did misread that. I thought they were saying there is now an "Allow access to ~/Documents?" prompt that stands for full disk access. I didn't remember seeing anything like that and I know it sounds ridiculous, but I guess it shows how low my expectations are for MacOS lately.

Apple tried the first thing first. That was App Sandbox, and you can grant access to specific folders through the open dialog (or better yet, use the open/save dialog to access individual files so that no broad permissions are needed at all). But nobody adopted App Sandbox unless forced to.

Negative access is what came next, applied to all apps. By default, an unsandboxed program can access everything except specific protected directories, and can request access to those directories too, e.g. Documents, Desktop, Photos, other apps' data.

Full Disk Access is the last resort that turns all of that off. Or, well, was.

It's also the only way to access Messages, which might be why certain AI apps request it.


Powerpoint/Word on Mac will pop up a slightly confusing "grant permission to specific folder" dialog if you try to rename the file you are working on. GIMP does not, but somehow needs to be specifically given access to the clipboard, which Powerpoint has by default.

At this point, permissions on Mac is in such a broken state it does not add anything of value. It is extremely common for apps to open up dialog boxes telling you exactly where to go in system settings to grant them access. Malware is just going to do the same, and users are trained to expect it.


Permissions on macOS work better than any other desktop OS by miles, it's not even close. But the GIMP is not a native Mac app and probably doesn't do things the way it needs to for proper integration. Use Pixelmator or some other Mac graphics app and things work fine.

Apps ask you to manage the system settings dialog only when they're requesting a very high level of system access and ship outside the app store. That describes a lot of apps that developers like to use (custom window managers etc), but not so many apps normal users will encounter.

When apps do uncover use cases that the permissions system isn't a good fit for, Apple are pretty responsive about adapting the system in backwards compatible ways. For example, AI apps wanting full disk access because they want to read the internal files of a messaging app - not ideal, so Apple have introduced a way to grant access to only specific apps from a modal popup triggered when the app tries to do it, not from the system settings app.

The way this is meant to work, by the way, is that the messaging app cooperates with the apps that want access to messages by exposing an app over XPC. AI apps are going offroad because AI companies in such a mad scramble to ship features and because of the complicated web of corporate relationships - does Meta really want to grant OpenAI access to people's WhatsApp messages on macOS or would they prefer to push users towards Meta Muse, etc. So AI apps bypass the genteel cooperative way and try to suck data out of the app's internal files directly. Apple is standing in the middle trying to mediate this stuff in a reasonable way, and I think they're doing a fine job.


If your desktop OS security solution only works for a small subset of your apps and in inconsistent ways, with no way for the user to trust the behaviour, you are not really doing it any better than any other system. You are only providing security theater.

Any sort of malware or malware-ish app (including software from a lot the large tech companies like Meta) will know exactly how to circumvent the protections, and exactly how to get the users consent in the process.


> By default, an unsandboxed program can access everything except specific protected directories, and can request access to those directories too, e.g. Documents, Desktop, Photos, other apps' data.

Not quite. Not every apps obeys this and I’m not even sure how. Adobe’s apps all insist on creating an “Adobe” folder in your iCloud-synced Documents folder regardless of file system permissions. You may deny access to every Adobe app under the sun, and they will still create those goddamned folders at launch. Apple’s solutions are not universally implemented and are most definitely not universally obeyed.

The day Apple gives me a way to simply say “fuck you Adobe, you won’t have access to my Documents folder anymore” is the day that these permission models have started to work properly, until then it’s all theatre.


Isn't the entire premise of the system being discussed here is that developers are forced to use it? So then why not just force developers to use the first thing you described?

It would break existing applications. Apple is hardly like Microsoft when it comes to back-compat, but they know enough to not break everything overnight: they can’t damage macOS’ credibility as a “real” desktop OS too much.

The main reason is that it’s a simple fact-of-life that software needs to be able to interact with other software, with or without permission (e.g. part of my day-job involves LoB integrations between software vendors that either went-out-of-business, despise each other, or live on vendor lock-in and offer no export/import functionality (looking at you, Progress, Clarion, and others…) - my work would be very difficult - but not impossible I’ll admit - if it ran on

(Fwiw, I always follow PLP when creating daemons, their service-principals, and FS ACLs, I’m always amazed at how everyone else (esp. multimillion-dollar ISVs) runs their MSSQL Express instances under SYSTEM, with ‘sa’ enabled, where every install uses the exact same single hardcoded password; often with LAN router port-forwarding on 1433 for “remote support” (don’t forget TeamViewer or Splashtop too). Sorry I’m ranting but the amount of security-incompetence still rampant in the LoB software industry today is flabbergasting)


> If you can't perceive it yourself, find someone who can.

Or just use your phone’s slo-mo feature to take a 240 FPS video (which then gets played back at 30 FPS). 60 Hz flicker becomes extremely visible. More than once I’ve noticed the strobe effect and then confirmed it this way.


That's a really good idea, thank you.

That looks pretty readable to me. Somewhat noisier than SwiftUI due to all the repeated "cui_", but not much. I've seen far worse.

I would be more concerned about the opposite problem: the fact that it's apparently reactive and thus inherits the same performance pitfalls as every other reactive UI framework. Retained-mode UI is much uglier (at least if you build the UI in code) but doesn't have these issues.

edit: actually it's worse than reactive, it's full immediate-mode render-everything-every-frame.


I'm observing the opposite phenomenon. I recently vibecoded (with supervision) a 25k-line Rust codebase that takes 15 seconds for every build. If I were writing that by hand, a 15-second build time would drive me crazy and I would stop working until I could find a good way to speed it up. With vibecoding it barely matters, because the interval between builds is much longer and the agent (Opus 5.5) spends most of its time thinking or running tests. Unlike me, the agent rarely makes the kind of trivial mistakes that require rebuilding multiple times in a row.

For larger codebases, things might be different.


Why GPT-5.6? Was this done before the release of GPT-6, or is there some difference in restrictions between the two?

Daybreak blue (for defensive action) was for a long time a specific model selection, basically Sol 5.6. Later it became a toggle available for any 5.6 model. Last time I checked it wasn’t available for 6+.

Daybreak red for offensive action is a better trained model of Sol 5.6 for offensive action.


I see. Interesting.

Personally I think LLMs did only produce garbage code. Not coincidentally, the same upgraded LLMs that have gotten so good at finding bugs are also much better at not writing bugs. Especially if you make them review their own work. They still have a hard time with code design, though.


I'm skeptical that this particular task would get blocked, though I could be wrong. But I agree that the moat is a problem in the general case.


I don't know about cryptography, but if you try anything in organic chemistry or biology, you virtually always get flagged and downgraded.

As this is going on, Anthropic has opened up their own biology lab.

> https://www.reuters.com/world/anthropic-quietly-sets-up-biol...

There is very much a situation where people in frontier labs have access to exceptional capabilities that are being kept from the rest of us, and they're using it to strip-mine technical and scientific problems.


I ran into this situation and was able to work around it by running a jailbroken Chinese model that turned out to be quite capable. I think it was Qwen.


Strip mining is a great phrase here, thank you


I can't take any credit for it. I believe Terence Tao coined it first. He also wrote a bit about it here: https://mathstodon.xyz/@tao/117237320796901560


Agree.-

They sure as hell ain't stripteasing :)


The strip teasing is all of their capability gating.


Come to think of it, indeed.-


> There is very much a situation where people in frontier labs have access to exceptional capabilities that are being kept from the rest of us, and they're using it to strip-mine technical and scientific problems.

If these capabilities were available to the public, someone somewhere would make a biological weapon or create a virus we've never seen before that could kill many millions of people.

Stanford researchers have demonstrated it’s possible [1].

[1]: "AI creates 16 new viruses from scratch, showing promise for drug resistance and drawing warnings about potential for misuse" - https://www.cnn.com/2026/08/06/health/ai-viruses-bacteriopha...


Biologists already addressed this idiocy a few weeks ago [1]. Making a biological weapon requires facilities that only some nation states - and comparable companies- have access to.

AI isn't going to magically assemble a DNA sequence for you and a computer design is not a guarantee of something that works.

AI restrictionists love to pretend that their toys will, with the power of intelligence, magically bypass fundamental limitations of reality to conjure up fantastical weapons of mass destruction if they throw enough computers at the task. Unfortunately for them, that is not how the world works.

[1] https://www.businessinsider.com/researchers-ai-apocalypse-sk...


We already have controls for that: we control access to the reagents and equipment needed. Yknow, like we always have.

Information is not dangerous. This is a ridiculous argument to me, but maybe my chemistry degree has me biased.


Not to mention that making an actual novel virus requires thousands of dollars of equipment, serious lab skills, and even then is so dangerous that the wannabe terrorist is more likely to infect and kill themselves than cause mass causalities.

It's the equivalent of making pipe bombs at home.

The instructions are out in the public, but that's not a problem because the idiots that make pipe bombs blow themselves up more often than they succeed in doing terrorism.


Broadly, the downgrading is still way too heavy-handed and clumsy. As of yesterday “How easy to crack is WPA2” downgraded to opus 4.8. Granted I could probably have phrased it differently and gotten a response, but the point is I shouldn’t have to. It can’t simultaneously be smart enough to reverse the ECDSA curve seed and not smart enough not to destroy the world by answering a trivial question with a scary word in it.


Fable 5.1 has downgraded me in the past week alone for the crimes of describing a coordinating server as "the C&C server" ([cyber]), for doing some statistical analysis on some toy Vigenère-ciphered text ([cyber]), and for asking it to help me find patterns in some semi-random text ([bio]), so it is fair to say I am not skeptical.


I have access to Claude's "defensive" cyber model and it appears to take no issue with the prompt, but I suspect that is because it uses today's publication as the basis for an answer.

In the past my attempts to work on crypto related problems have been met with needing the higher tier of "offensive" access.


It doesn't matter what is blocked today, it matters what is blocked tomorrow.

If we allow these providers to promote regulatory capture they become the arbiter of what anyone is "allowed" to do


I think when the open source models catch up (likely soon) we'll long for the days when companies could control it like this. I could easily see a teenager with openclaw and a single prompt causing millions of dollars in damage. Multiply that by the number of stupid people in the world and we have a problem.


At minimum we don't have access to "Claude thoughts" like the ones the article highlighted, right? That's part of the hidden reasoning?


Correct.


Presumably yes. It is safe to assume that algorithms running the model changed, the training procedures changed, and the scale changed. Unfortunately the AI companies refuse to disclose the details of any of this, which is a harm to human knowledge (hopefully a temporary one). Yet the results speak for themselves.

For me it's only been in the last month with Opus 5.5 that I've been seeing code quality I actually respect (though this is with substantial human effort in the loop). Though this experience is from work where I don't have access to non-Anthropic models, nor Fable.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: