Stories
Slash Boxes
Comments

SoylentNews is people

posted by hubie on Thursday October 08, @08:48AM   Printer-friendly

Meta says FDA isn't sufficient to Muse reading messages. Apple begs to differ.

Apple says it is changing its macOS privacy settings to stop third-party app developers from misusing them to access message histories.

Friday's announcement comes two weeks after tech columnist Jason Aten said that Meta's new general-purpose AI agent Muse sent him an unsolicited notification referencing a thread between him and a co-worker over Apple Messages. Aten said he never granted Muse permissions to read his messages and had assumed they were off-limits. Social media last week blew up with masses of people who agreed and said the incident showed that AI assistants given access to calendars, emails, messages, shopping accounts, and other resources are akin to a skill saw or other power tool. While potentially useful, they can do real damage if not used carefully.

Meta CTO David Singleton joined the fray with a rebuttal that appeared solid. For Muse to access Apple Messages, a user must manually give it two privileges. One is full-disk access, a macOS system-level permission. The other is to enable a Messages connector setting in Muse.

"The Messages integration in the Muse Mac app is opt in," Singleton said. "Your Muse can only read Messages content if macOS system-level Full Disk Access is granted and the Messages connector is enabled."

Singleton's implication was clear. Muse could have read Aten's Messages communications only if he had enabled both settings, and if so, the columnist had only himself—and certainly not Meta—to blame.

Earlier this week, I spoke to macOS security expert Patrick Wardle, who questioned Singleton's denial. His reasoning: "From a technical point of view, with FDA (full-disk access), any (non-root file), is readable, browsing history, browser cookies, chats, etc etc etc." I asked Meta how Muse couldn't read messages when the app had full disk access, while every other app with that privilege could. Meta PR's only response was to requote Singleton saying: "The Messages integration in the Muse Mac App is opt-in. Your Muse can only read Messages content if macOS system-level Full Disk Access is granted and the Messages connector is enabled."

Now, Apple is setting the record straight. In explaining why it was going to make changes to the FDA permission setting, the company wrote:

Some developers are using Full Disk Access in ways that could put users at risk, exposing everything on their systems—including files, mail, messages, and even browsing history—without users' full knowledge and understanding. For communication apps, this can also compromise the privacy of the people users are communicating with.

[...] Taken together, the events suggest that Muse may not be worthy of the extraordinary access it must have to work as billed by Meta. People who use the assistant should configure permissions carefully, though as Aten's experience suggests, that precaution only goes so far.


Original Submission

This discussion was created by hubie (1068) for logged-in users only, but now has been archived. No new comments can be posted.
Display Options Threshold/Breakthrough Mark All as Read Mark All as Unread
The Fine Print: The following comments are owned by whoever posted them. We are not responsible for them in any way.
(1)
  • (Score: 1, Interesting) by Anonymous Coward on Thursday October 08, @09:13AM (4 children)

    by Anonymous Coward on Thursday October 08, @09:13AM (#1456021)

    My AI Agent is sandboxed in a VM and only has web access via a web proxy server and access to the AI Stuff via haproxy.

    In theory it could still escape by a VM escape exploit ( https://en.wikipedia.org/wiki/Virtual_machine_escape [wikipedia.org] ). But it probably would also need a local privilege escalation exploit (the agent is running as a normal user account) so I think it's still safer than the fake[1] "sandboxes" that Anthropic, OpenAI, etc have been using that don't even need VM escape exploits to break out of.

    [1] Call me cynical but I think they intentionally left such exploitable gaps in their sandboxes. Because "Hey everybody, our AI managed to hack (only) our own VMs!" probably doesn't even get an "Oh no! Anyway..." response, in contrast "Our AI managed to hack other people's stuff" gets a lot more publicity. They've got enough money and resources to sandbox stuff far more than I'm doing - real airgaps, local repos for more security and better reproducibility, etc. It's more scientific to have local known repos, so you can repeat the experiment/test again, with the same AI seed, but maybe change something just a bit and see what happens this time around.

    • (Score: 4, Informative) by Frosty Piss on Thursday October 08, @04:54PM (3 children)

      by Frosty Piss (4971) on Thursday October 08, @04:54PM (#1456056)

      If you had kept up to date on the dark side of AI "agents", you would know that breaking out of a sandbox is a trivial pursuit to them.

      • (Score: 1, Informative) by Anonymous Coward on Friday October 09, @03:32AM (2 children)

        by Anonymous Coward on Friday October 09, @03:32AM (#1456083)

        OK lets see your citations that breaking out of a sandbox is that trivial for them (given an up to date sandbox environment).

        From what I've seen:
        For the HuggingFace incident the OpenAI agents needed to use an exploit in Jfrog (which shouldn't have been given access to the internet).

        In the Meta, Google and Anthropic agents "escape" was allegedly due to misconfiguration which gave the AIs direct access to the internet:

        https://www.reuters.com/business/gemini-hacked-three-companies-first-known-breakout-by-google-ai-wsj-reports-2026-09-18/ [reuters.com]

        Gemini model accessed the internet and hacked other companies during a test of its cybersecurity capabilities, ​the first known example of the company's AI systems autonomously committing ‌such an act.
        The hacks occurred in May during a cybersecurity test conducted by Irregular, an independent company that conducts cybersecurity evaluations.

        https://www.engadget.com/2231446/meta-ai-model-hacked-third-party-irregular/ [engadget.com]

        Stone said the model was able to access the internet due to a misconfiguration in the testing environment by the company's evaluation partner Irregular. After gaining access to the internet, it then exploited a security vulnerability in a third-party service, "in a matter similar to previously reported instances with other companies."

        It was also due to a misconfiguration by Irregular that Anthropic's models were able to leave their testing environment and hack into three organizations. When the company announced the incidents, Anthropic laid the blame at Irregular's feet. OpenAI also reported an incident separate to the Hugging Face hack, where its models were also able to access the internet because of, you guessed it, the same testing partner.

        Sounds more like publicity stunts to me.

        Even their own AIs would be able to design and configure better sandbox environments. Stuff like frozen mirrors/repos with no internet access for better test reproducibility.

        • (Score: 3, Insightful) by arslan on Friday October 09, @03:44AM (1 child)

          by arslan (3462) on Friday October 09, @03:44AM (#1456085)

          That's one way to look at it. The other more plausible explanation is they actually did misconfigure it much like many of the data breach events we've seen in the past, non-AI related, where organizations fail under the weight of their own bureaucracy and lead to misconfigurations that creates news events. And both can be true as well!

          I'm not questioning your skillz in sandboxing, but I would imagine Apple is trying to address a majority of their user base which I can only imagine is not as capable as yourself - they'd know for sure of course, I'm just speculating.

          • (Score: 0) by Anonymous Coward on Monday October 12, @01:17AM

            by Anonymous Coward on Monday October 12, @01:17AM (#1456272)

            The other more plausible explanation is they actually did misconfigure it much like many of the data breach events we've seen in the past, non-AI related, where organizations fail under the weight of their own bureaucracy and lead to misconfigurations that creates news events.

            I still find that not plausible because I tried asking Claude Opus to design a good sandbox, and it came up with a better safer one than they used. The targets were isolated. The hypervisor host itself would be running on dedicated bare metal hosts used only for the sandbox[1].

            As for Apple, the real issue is their user base probably has significant percentage who actually want AI to read/change their data.

            There's no sandbox that works for that. If you want your AI to be able to read your files, it's going to be able to read your secrets if they are saved among those files. If you want the AI to be able to delete your emails, your AI is going to be able to delete your emails by mistake.

            [1]

            Design a sandbox environment for an AI agent that's to do penetration tests. There needs to be a repository in case the agent needs to download tools etc. The agent should NEVER be able to escape the sandbox environment.

            Sandbox design for an AI penetration-testing agent

            Treat the agent as untrusted code that is actively trying to get out. It might do this through its own misbehaviour, or because a target system feeds it a prompt injection such as a malicious banner, web page or file. No single control can promise it will never escape, so this design stacks independent layers. An escape would need several unrelated failures at once, and each layer is enforced from outside the agent's reach.

            1. Architecture
            flowchart LR
                    subgraph MGMT["Management zone (trusted)"]
                            ORCH["Orchestrator / policy engine"]
                            LLM["LLM API client"]
                            LOG["Log & PCAP store (WORM)"]
                            HUM["Human approver / kill switch"]
                    end

                    subgraph HOST["Dedicated sandbox host (bare metal, no other workloads)"]
                            subgraph VM["Agent microVM (Firecracker/Kata) - ephemeral"]
                                    AGENT["Exec runner + pentest toolchain"]
                            end
                            FW["Hypervisor-level firewall (nftables on host, outside guest)"]
                    end

                    subgraph RANGE["Target range (isolated VLAN/VRF)"]
                            T1["Target VMs"]
                    end

                    subgraph REPO["Internal repository (read-only to agent)"]
                            MIR["Curated mirrors: apt, PyPI, Go, GitHub releases, Docker images"]
                    end

                    INET(("Internet"))

                    ORCH -- "vsock: run_command() only" --> AGENT
                    LLM --> ORCH
                    ORCH --> LOG
                    HUM --> ORCH
                    AGENT --> FW
                    FW -- "allow: in-scope CIDRs" --> RANGE
                    FW -- "allow: HTTPS GET to mirror only" --> MIR
                    FW -. "deny all else" .-x INET
                    INET -. "offline / one-way ingestion, scanned & signed" .-> MIR

            The main rule is that the model's "brain" never lives inside the sandbox. The orchestrator runs in the management zone and calls the LLM API from there. It sends commands into the VM over a narrow channel and receives output back. The sandbox therefore needs no internet access, no API keys and no route to the management zone.

            2. Isolation layers (defence in depth)
            Layer Control WhyPhysical Dedicated bare-metal host(s) used only for the sandbox. No shared tenancy with production or management systems. A hypervisor escape then lands on a host with nothing of value on it.
            Virtualisation One microVM per engagement (Firecracker or Kata). Don't rely on plain Docker. Disable nested virtualisation and keep virtio devices to a minimum. A kernel boundary is much stronger than a container namespace boundary.
            Guest hardening Read-only root filesystem. A writable /work scratch volume only. No host mounts and no Docker socket. Seccomp and AppArmor profiles. CPU, RAM, disk and PID limits. Shrinks the attack surface and blocks resource exhaustion.
            Network Default-deny firewall enforced on the host or physical switch, never inside the guest. Separate VLAN/VRF with no route to management subnets. Block 169.254.169.254 and other cloud metadata endpoints. The agent has root in its VM and can rewrite any rules inside it.
            Control channel vsock or a serial-style RPC exposing only run_command, read_file and write_file within /work. Strict schema, size limits and timeouts. Orchestrator code is minimal and audited. This channel is the only path to the trusted side, so keep it as small as possible.
            Secrets No credentials in the VM except engagement-scoped test credentials. LLM keys and the repo signing keys stay outside. Nothing worth stealing if the VM is compromised.
            Lifecycle Rebuild every VM from a signed golden image. Destroy it at the end of the engagement or on any anomaly. Never reuse a VM across engagements. Persistence and cross-engagement contamination become impossible.
            3. Tool repository

            The agent can download tools, but only from what you have already vetted.

            Repository contents:

            Mirrors of Kali/Debian apt (aptly or reprepro), PyPI (devpi or Nexus), Go module proxy (Athens), Docker/OCI images (Harbor), and a GitHub release cache. Artifactory or Nexus can cover most of these in one place.
            Exploit and wordlist collections such as SecLists and the Metasploit modules.

            Getting packages in (one-way):

            A separate ingestion host with internet access pulls updates on a schedule or on request.
            Every package is checked for malware (AV and YARA), checksums and signatures are verified, and an SBOM is generated.
            Approved artefacts are re-signed with your internal key and pushed across a data diode, or a firewall rule that only allows ingestion → repo.
            The repository has no outbound route anywhere.

            Agent access:

            The firewall allows HTTPS from the agent to the repo on one port and nothing else.
            The agent gets read-only credentials, and package managers inside the VM trust only your internal signing key.
            Rate-limit downloads and log every one.

            Missing tools: the agent calls a request_tool(name, reason) function. The request goes to a human queue, and an approved tool goes through the same ingestion pipeline. The agent never fetches from the internet itself.

            4. Scope enforcement

            The allowed target list is defined per engagement in the orchestrator. It is compiled into host-level firewall rules when the VM is created, so the agent cannot widen its own scope.

            Lab or range work: targets sit on an isolated VLAN with no route out.
            Live client engagements: the agent's traffic goes through a dedicated egress gateway. That gateway allows only the in-scope IPs, ports and time window from the signed rules of engagement, and everything else is dropped and alerted. This mode is the riskiest, so add rate limits and human approval for destructive actions such as exploitation, DoS-prone modules and credential spraying.
            5. Monitoring and kill switch
            Full packet capture on a host-side tap, plus logs of every command and its output. Store these in write-once (WORM) storage outside the sandbox.
            Tripwires:
            Traffic to denied destinations.
            DNS queries for external names.
            Probes against the vsock or orchestrator.
            Hypervisor or kernel exploit signatures.
            Unusual syscalls (Falco or eBPF on the host).
            Canary tokens placed in the VM.
            Automatic kill: any tripwire freezes the VM. Snapshot it for forensics, then destroy it.
            Human kill switch: a physical or out-of-band control that cuts power or the network to the sandbox host.
            6. Results out

            Findings leave only through the orchestrator as structured data, such as JSON findings and specific files under a size cap. Scan them before they reach analysts. Treat everything coming out as potentially malicious, because target data can carry payloads or prompt injections aimed at whoever reads it next.

            7. Keeping the sandbox trustworthy
            Patch the hypervisor and host kernel promptly; this is your most critical dependency.
            Red-team the sandbox itself regularly: give a test agent an explicit "escape" goal and fix whatever it finds.
            Run independent reviews of the orchestrator's control-channel code and the firewall rule generation.
            Keep management access to the sandbox host on a separate admin network with MFA and no shared credentials.

            Minimal viable version: a dedicated host running Firecracker VMs, nftables rules on the host, a Nexus mirror fed through a scanned one-way pipeline, the LLM kept outside and commanding the VM over vsock, ephemeral VMs, and full logging with a kill switch.

  • (Score: 1, Troll) by Anonymous Coward on Thursday October 08, @12:04PM (5 children)

    by Anonymous Coward on Thursday October 08, @12:04PM (#1456029)

    smart move apple. they understand their product sells to the mentally retarded and have consequently turned their operating system into a padded cell so the customers can't hurt themselves.

    • (Score: 1, Interesting) by Anonymous Coward on Thursday October 08, @12:18PM

      by Anonymous Coward on Thursday October 08, @12:18PM (#1456032)

      iSell: expensive products to dumb poor people who want to look smart.

    • (Score: 4, Interesting) by VLM on Thursday October 08, @04:07PM (1 child)

      by VLM (445) on Thursday October 08, @04:07PM (#1456053)

      they understand their product sells to the mentally retarded

      Mentally ill, actually. Seems to correlate with depression symptoms and anxiety symptoms.

      https://pubmed.ncbi.nlm.nih.gov/14723468/ [nih.gov] "... Macintosh users had significantly greater anxiety about computers than PC users...." etc.

      The cellphone data can be weird, anxiety and depression symptoms seem to result in "normal" use levels, but the symptoms do seem to correlate with blaming the phones for the illness symptoms. The study below didn't report a difference between mental illness levels of iphone vs android users although subjectively I've seen iphone users are much whackier on average.

      https://researchportal.bath.ac.uk/en/publications/quantifying-smartphone-use-choice-of-measurement-impacts-relation/ [bath.ac.uk]

      I would propose that installing a program to read your messages and then having a meltie about it reading your messages is not exactly a display of mental stability or lack of paranoia symptoms. Like breathlessly reporting a new conspiracy theory that installing Minecraft causes blocky graphics on your computer but its early days and more research is needed.

      • (Score: -1, Redundant) by Anonymous Coward on Friday October 09, @03:35AM

        by Anonymous Coward on Friday October 09, @03:35AM (#1456084)
        More of them are gay too.
    • (Score: 1, Informative) by Anonymous Coward on Thursday October 08, @06:04PM

      by Anonymous Coward on Thursday October 08, @06:04PM (#1456059)

      How is this different from you as a linux user running a program that has the same read access as your username?

    • (Score: 1, Insightful) by Anonymous Coward on Friday October 09, @03:48AM

      by Anonymous Coward on Friday October 09, @03:48AM (#1456089)

      As opposed to Microsoft that also treats their customers as such and consequently turned their operating system into a pothole of hurt so their customers can hurt themselves?

      Or linux where it is all about the developers and their moral technical high-ground and not at all about potential non-nerd users?

      Which cave have you been living in?

  • (Score: 2, Interesting) by JoeMerchant on Thursday October 08, @12:14PM (4 children)

    by JoeMerchant (3937) on Thursday October 08, @12:14PM (#1456031) Journal

    Why call it Full disk access if you can't access the full disk?

    This sounds like a job for a ufw style permission chain: put your highest priority rules first, when a permission is sought the first match defines it.

    Private messages: deny
    Emails: deny
    Full disk access: grant

    Once you hit Full disk access, that's the end (unless you are including cloud access in this scheme)

    Note, if you make a habit of screen shorting messages, those would be accessible under photos. If you save email attachments those would be accessible under downloads unless you specifically deny photos and/or downloads access.

    What I want full search access to is the semantic content of the ambient microphone captures.

    --
    🌻🌻🌻🌻✌️ [google.com]
    • (Score: 2) by arslan on Friday October 09, @03:51AM (3 children)

      by arslan (3462) on Friday October 09, @03:51AM (#1456091)

      Lol, I concur. They should rename it to FDAA, Full Disk Access Attempt, which is summarily followed by "Access Denied"!!! The sheer user experience impact opportunity here is brilliant! Steve Jobs must be rolling in his grave, though he should be used to it by now!

      • (Score: 2) by JoeMerchant on Friday October 09, @04:23AM (2 children)

        by JoeMerchant (3937) on Friday October 09, @04:23AM (#1456092) Journal

        Still running 25-30% profit margin, net profits (inflation adjusted) are about triple when he died - I think he would have been very satisfied, except for the whole not being in total control thing.

        --
        🌻🌻🌻🌻✌️ [google.com]
        • (Score: 2) by arslan on Friday October 09, @04:50AM (1 child)

          by arslan (3462) on Friday October 09, @04:50AM (#1456095)

          I could be wrong, but he didn't strike me as a profit first guy but more like changing the world or realizing his vision - depend on how you see it, his way (queue the control thing). Of course, like everyone else, profit is still quite desirable, but I didn't think it was his primary objective above all else unlike his contemporary counterpart Bill.

          • (Score: 2) by JoeMerchant on Friday October 09, @01:42PM

            by JoeMerchant (3937) on Friday October 09, @01:42PM (#1456105) Journal

            Well, we live in a basically capitalist (with heavy distortions) society. So, without capital you have no power to change the world - any successful changing of the world generally requires heavy concentration of capital.

            My impression of Jobs, especially as contrasted with Woz, was that he had his eye on the capital first - perhaps as a means to make his black turtleneck fashion statements to an audience that paid attention - but capital first for Steve, not so much as a desire but as a prerequisite imparitive, ever present. I never saw a significant Jobs philanthropy play, his enthusiasm was always part of a massive cash generation play - not that they all succeeded, but they all seemed to be aimed at that target - especially the cheap computers for school "charity" play. Those computers were intentionally nerfed, short lived, not something that would ever cannibalize a commercial sale, basically purpose built brand promotional devices.

            Gates just didn't consider image, and when you're on track to be the richest man in the world and don't give a damn about image, people aren't going to praise your kind heart. At least Gates (and Bezos) married philanthropists and gave them significant control of their hoards.

            --
            🌻🌻🌻🌻✌️ [google.com]
  • (Score: 4, Interesting) by Rich on Thursday October 08, @04:53PM

    by Rich (945) on Thursday October 08, @04:53PM (#1456055) Journal

    Think about Apple and Meta what you will, but in this case the journo seems to be completely clueless, and both FDA and Message Access for the plugin seem to have been explicitly granted. Of course the guy can't remember activating something when he has no clue what it all means, maybe following some slop video instructions on Youtube.

    Once you have assigned FDA to an application, that application can access everything within UNIX permissible rights for a given user, and knowing the application file formats, can therefore access whatever messages or other data they store. With sandboxing active for an app, it will be allowed to access files to be opened by user action (via drag & drop, or file-open dialog) and beyond that adhere to a not clearly communicated ruleset. E.g. a "normal" application trying to access something within the "Documents" folder will throw a permission request for that folder, whereas it might not ask for a non-blessed "My Docs" folder right next to it.

    In general, cross access within user rights was considered desirable, because it allows many small tools to interact to build a workflow, where the power user had to be trusted. But I fear with the amount of complete retards now letting AI agents on the loose, local zero trust will have to be maintained to an extent where such people are too stupid to deactivate it. Which is exactly what Apple plans now, to the annoyance of every power user.

    There's a special case, by the way, which is "Terminal.app". When Terminal has FDA, it has free reign for everything it executes in its zsh. Actually my Terminal.app on a Tahoe Mac on the right does _not_ have FDA enabled right now, but it can list both ~/Documents (which I may have granted a while ago, I don't remember) and /dev.

  • (Score: 2, Flamebait) by looorg on Thursday October 08, @05:18PM (1 child)

    by looorg (578) on Thursday October 08, @05:18PM (#1456057)

    Granting things that doesn't need it full disk access seems beyond stupid. Is this some kind of residue from when MacOS was not MacOSX based and was just a stupid single user system without a care in the world? Or are the developers so lazy and stupid that they can't develop without full disk access for all their trivial applications? They make it sound like almost every single phone app that wants access to everything just cause it's easier then to be specific. Better to grab it all then ask for permissions.

    • (Score: 2) by Rich on Thursday October 08, @07:07PM

      by Rich (945) on Thursday October 08, @07:07PM (#1456062) Journal

      There are two client types for FDA.

      One is ancient legacy apps that indeed might have come from classic Mac OS, or even OS X from before there was a sandbox. Up to about the highly praised Snow Leopard 10.6, which had to make do with what UNIX provided out of the box. Apple later added SIP (super-status for root-immutable system files) and the sandbox, and even later the sealed system volume. If no one had the resources to revise the software to the sandbox' needs, FDA is the workaround. By extension, lazy phone app developers would also fall into that category.

      I work on a legacy application that expects a defined data set in the "Documents" folder. Actually it was my idea to put it there three decades ago, still on Classic Mac OS. Thousands of worldwide installations are trained to expect the user-specific application data exactly there, exactly named like the product in that known location, according with all the documentation printed for this. Nowadays, such a data set might be relocated to "Application Support" - but users are used to shuffle around that app specific folder, and they would not find it without navigating to the hidden ~/Library first (where they then could do more harm than good). Also, all documentation in dozens of languages would have to be re-done and re-printed. We don't have to use FDA, but we have to nod off the "Documents" access once after installation (or, in my case, every compile that generates a new binary), because we don't explicitly open the work data with a file select box.

      The other client type is anything disk tool. Backup, usage analysis, such stuff. They really need it.

(1)