The attacker was a takedown extortionist: someone who demands ransom from group owners in exchange for not targeting their communities. These extortionists use automated accounts to plant illegal content in public groups and then report it directly to Apple, attempting to trigger the removal of legitimate communities whose owners refused to pay them.

Welcome to post 58 in the R⁴ series.

R and the CRAN repositories maintain a very high level of what we might call “quality assurrance” by requiring that newly-added code does not break any existing dependencies. This is frequently called a “reverse-dependency check”. For any given CRAN package one can quickly determine it reverse dependencies. Calling tools::package_dependencies(pkgName, reverse=TRUE) will for a scalar or vector-valued argument return a named list with the reverse dependencies. It is then a matter of looping over this list. There are helper functions in base R as well as in contributed packages on and off CRAN. I also wrote my own with package prrd which, while possibly a wee bit specialised and under-documented has served me well to check on Rcpp and related packages which can indeed have a large number of reverse dependencies.

I recently looked into one of these contributed runner packages, and while I will refrain from naming its implementation language let me just mention that the term “cargo cult” may be a real thing here. What go me interested in this was the fact that if one has a simple-to-use runner then the fact that r2u makes it “fast, easy, reliable: pick all three” (to borrow its slogan) to deal with actual depencies if Ubuntu has indeed been selected as the host. We will maintain the position that if you can in fact integrate with the system-wide package management then any alternative per-repo package management approach not doing so will likely be dominated by an approach that does integrate with the system facilities. Which is what precisely what r2u does, and offers. And why it is used enough to by now have shipped eighty eight million binary packages. So I tested it for the reverse-dependency check task.

What I learned by looking into the (much more complicated) runner was that it at the end of the day it hands the actual task of running the reverse dependecies off to a helper function rev_check that is part of the xfun package by Yuhui. I quickly found that besides xfun we would also need its suggested dependency tinytex which in turn would error unless the tlmgr binary was present. So as the sole requirement (on an Ubuntu system with r2u) turns out to be

$ apt install r-cran-xfun r-cran-tinytex texlive-base

where we do it all in one apt call (as root in the container). (Given r2u we could also call install.packages(c("xfun","tinytext")) followed by apt install texlive-base but it is simpler for this setup step to be just one call).

With that we are basically done. I did this (twice) using a rocker/r2u container with r2u preinstalled, mounting a local work and scrap directory for the container. In it we expand the package to be tested (i.e. tar xaf pkgName_*tar.gz for a given source package pkgName from CRAN) and then just call with the package name and expanded direcrtory. I.e. I used this call to test my package AsioHeaders (which has just three reverse dependencies) to both name it and to point to the expanded source directory created for this purposed:

> system.time( res <- xfun::rev_check("AsioHeaders", src="AsioHeaders") )
## ... earlier output omitted for brevity here ...
   user  system elapsed 
 35.732   3.333 149.683 
> res
   httpgd ipaddress websocket 
        0         0         0 
> 

and about a good two minutes later I would get the timing result and the summary in variable res. As I checked the current CRAN version, the check was as expected free of concerns or issues.

To support this, r2u did indeed go off and install about sixty seven binary packages (and the total includes all binary dependencies fully resolved) delivering on the ‘just works’ promise by the r2u documentation.

As another check, I did the same for RcppAnnoy which has seven reverse dependencies and needed about two hundred CRAN packages to be installed. The full test took just over four minutes with the timing function reporting some nice gains from parallelisation as total user compute time was on the order of just under eight minutes. Again, test results were clean and free of worries as expected:

> system.time( res <- xfun::rev_check("RcppAnnoy", src="RcppAnnoy") )
## ... earlier output omitted for brevity here ...
   user  system elapsed 
471.765 378.220 266.855 
> res
   bbknnR  bigANNOY  blocking     scDHA    Seurat      uwot VectrixDB 
        0         0         0         0         0         0         0 
> 

Overall this was a rather useful quick excursion as it demonstrates that - existing functions can be used to orchestrate a reverse dependency check - with ‘reasonable’ dependency scale we can do this on a single machine quite easily taking advantage of parallel computing on multi-core machines - using r2u gives us fast, easy, reliable package installation making testing of packages we might not otherwise use or know a breeze - doing this in an ephemeral Docker container facilitates easy build-up of required resources and leaves no side effects behind which might affect our normal development environment

This post by Dirk Eddelbuettel originated on his Thinking inside the box blog. If you like this or other open-source work I do, you can now sponsor me at GitHub.

Now generally available

• Cisco Secure Access is an advanced enterprise browser integration for Edge for Business. As a single integrated connector, it brings not only secure access but also unified data protection and AI guardrails directly into the browser.
• Tanium Connector for Microsoft Edge for Business delivers real-time browser telemetry into Tanium Autonomous IT Platform, helping organizations gain deeper visibility into browser-based threats, close critical security blind spots, improve compliance validation, and strengthen overall endpoint security across the enterprise.
• Clever Classroom MFA TrustPass helps streamline logins in education environments by recognizing trusted devices in Edge for Business—reducing unnecessary MFA prompts while maintaining strong security.

Bringing it all together

We briefly removed Telegram from the App Store after our review found content that violates our strict guidelines prohibiting child sexual abuse material. The app was subsequently restored after the developer promptly removed the content and banned the user who posted it.

After seeing a video a few months ago demonstrating how a proprietary CAM solution uses machine learning and large language models to automatically generate CNC instructions, and successfully testing it on a real CNC, I began wondering if the same could be achieved with free software. I still do not know the answer, but I may be getting closer to finding out. Two weeks ago, I came across the video "I Connected Claude AI to FreeCAD (And It Models Parts Like an Engineer)" by Make Form, which introduced me to the FreeCAD MCP project. Even though the video creator apparently believes it is acceptable to download and run random binaries from the Internet on a local machine (his setup uses UCX), I do not. I would probably have left the project alone entirely if I had not noticed that all of its dependencies are already available in Debian. This significantly boosted my motivation, so I set out to test it using packages built from source on Debian rather than relying on untrusted binaries.

The first hurdle was that the MCP SDK for Python was not present on my Debian Forky test machine. I initially believed it was missing from Debian altogether, but it has been available in Debian Unstable for about a month and is only absent from Forky because some automated tests fail on architectures like riscv64 and s390. Fortunately, backporting it was straightforward using `apt-get source -b python3-mcp`. The next hurdle involved an outdated version of the Validators Python library. Since I am a member of Debian's Python team, which maintains this package, updating it to a sufficient version for FreeCAD MCP was relatively easy. I could not upgrade to the latest upstream release due to a new dependency on an Ethereum-related library, so I settled on a 2024 version.

With those dependencies in place, I proceeded to create a Debian package for FreeCAD MCP. I had previously submitted a request for packaging of FreeCAD MCP to gauge interest while deciding whether to prioritize maintaining it myself. Because salsa.debian.org blocks access from Tor users like myself, I published my draft packaging scripts in a Git repository on Codeberg as the Debian FreeCAD MCP project and got it working with the FreeCAD 1.1 version in Forky. I initially struggled with the button controls for the MCP feature, which led me to submit a pull request titled "Fixed startup sync of checkable toolbar buttons" proposing a fix. Once this confusion was resolved and the MCP setup was enabled via the GUI, I was able to run FreeCAD completely headless using `xvfb-run` on a machine without an X server to generate models. I am using a private LLM service running the Debian package of llama.cpp with the Qwen 3.6 model downloaded from Hugging Face, configured with a maximum context window of 105k tokens. I also tested the Bonsai model on my test laptop; initially, its context window was too small (8k and 16k could not accommodate the FreeCAD MCP instructions), but even after increasing it to 32k, it proved useless for generating FreeCAD models so far. I've used Claw Code, Aider and Open Code with my server so far, and for this test I ended up with OpenCode because it was easy to set up to use an MCP. Because none of my LLM services are set up to be multimodal (capable of processing both text and images in this case), I configured the MCP to return only textual feedback from FreeCAD. I am unsure if this is a major limitation, though I suspect it might be.

My testing experience remains limited, with no clear successes yet. Part of the issue likely stems from my ability to provide effective instructions for modeling 3D objects (I am relatively new to FreeCAD, English is not my first language, and I lack a precise vocabulary for describing construction features to an LLM). Nevertheless, the LLM has demonstrated the capacity to create 3D models in FreeCAD. In one of my first tests, I asked it to generate a cube and then produce CAM/G-code instructions for a CNC machine. It did output G-code (which remains untested), but I was surprised to find that it bypassed FreeCAD's built-in CAM module entirely and instead generated an external Python script to produce the code. This was not quite what I intended, though my instructions were probably unclear. The Qwen model with OpenCode seems to strongly prefer programming directly; it frequently executes Python snippets inside FreeCAD to achieve its goals rather than using the standard sketch-and-extrude workflow I am accustomed to. In another test, I asked the LLM to create a parameterized pipe assembly to see which of FreeCAD's parametric tools it would choose, but found no evidence of traditional parametric features in the output. When prompted, the LLM explained that the parameters were embedded directly in the Python script used to generate the model, rather than in native FreeCAD features. With more explicit instructions, it eventually created a FreeCAD spreadsheet to manage the parameters. The resulting model looked much closer to my expectations and could have been useful with further refinement. My so far last experiment was less successful: I asked it to design a pipe clamp, but the LLM repeatedly failed to position the clamping screws in a way that would actually secure the brackets around the pipe. It is unclear whether this limitation lies with the model, my prompt, or other factors.

Based on my testing so far, I am uncertain whether FreeCAD MCP is merely a fun toy or a genuinely useful tool. I will only commit time to maintaining it in Debian if it proves to be practically valuable. I would welcome feedback from anyone who has experience with the project, preferably via the original request-for-packaging mailing list thread. Alternatively, I am available in the FreeCAD and Debian AI IRC channels for further discussion.

As usual, if you use Bitcoin and wish to support my activities, please send donations to 15oWEoG9dUPovwmUL9KWAnYRtNJEkP1u1b.

• Copy text or other supported content on their iPhone and paste it directly on their Windows PC, and vice versa.


• Perform common productivity workflows without needing to foreground an app or take explicit repeated actions to initiate a transfer.


• Experience clipboard synchronization as a lightweight, continuous capability rather than a manual, app-driven operation.

To adopt this solution, you will need to adopt AccessorySetupKit in order for the user to consent to the necessary permissions allowing the pasteboard to be shared with your accessory. This solution will require the user to perform a one-time consent to sharing the pasteboard with each paired accessory that will receive this functionality.

• Round One, Ted Energy - Contestants have to introduce themselves in character and have a two-minute dance break. Judging will be based on optimism, posture, and presence, along with costume commitment and participation level.


• Round Two, Head Coach - Semifinalists will participate in a trivia game and will be judged on the level to which each participant embraces the Ted Lasso character. The round will be judged on trivia results, costume commitment, mustache authenticity (Apple says 'stick-ons OK; real grows extra credit'), and The Believe Factor (something unique and special).


• Round Three, Ted Talk - Each remaining contestant will present a 30- to 45-second pep talk to the crowd. It will be judged on how moved the crowd is, costume commitment, and Ted Energy.

Today I stumbled over some behavior that I found quite surprising:

$ ipython3 -c 'import httpx;print(httpx.URL("https://example.com/foo/bar/../../baz"))'
https://example.com/baz

Even more surprising that behavior is actually standards-compliant, even mandated by RFC 3986.

The underlying motivation is relative reverences. If some resource reachable by "https://example.com/foo/bar" references another resource relatively as "../../baz" then this is of course the intended result.

Getting from this problem to what RFC 3986 suggests might be surprising in the result, but somewhat understandable if you look at the consequences of that problem:

Giving the path components ".." (and ".") special meaning at the start of the relative reference means that if you allowed them in absolute URLs those would be impossible (or at least very convoluted) to address as relative URLs.

So RFC 3986 describes a way to handle them everywhere: Just join the path of the base URL and the path of the relative reference and normalize the result. Or normalize the absolute on either side if only that is to be taken. This makes things very convenient: Multiple reference URLs can just be joined without special handling for relative references starting with dots, making writing applications handling them easier. Programmers don't have to care how to handle relative references and can just join everything in whatever way they want.

For maximum elegance there is still some corner case left: What happens if an absolute URL has a path starting with double-dot components? Or an relative path starting with more of them then the base URL's path has components. You just ignore them:

$ ipython3 -c 'import httpx;print(httpx.URL("https://example.com/../../baz"))'
https://example.com/baz

With that last point every URL is valid and has well-defined meaning. Handling relative references and relative paths is very easy and convenient.

So this shows a high regard for simplicity, elegance and convenience. And a total and uncompromising disregard of security.

After all the most convenient it is for an attacker; If they are allowed to supply a path component for a request a system does in their behalf, then they can easily escape anything they were supposed to be limited to. The ignoring of dots at the start means they don't even have to know exactly how deep their request is:

$ python3 -c 'import httpx;print(httpx.URL("https://example.com/public/api/public/resources/harmless/../../../../../../../../../internal/data"))'
https://example.com/internal/data

So even if the resource server securely handles request (unless you consider not having any way to lower your permissions for one request to a specific subset), your fully RFC conforming client library will already request the permission they should not have permission for. Even worse dots are usually not characters you can easily forbid so once slashes are to be allowed things get complicated.

There also would have been a simple, elegant and secure way: Consider every path element ".." or "." in an (absolute) URL an error. Define a reference resolution that allows the relative reference to only start with "./" or one or multiple "../" and consider every appearance of a dot or two dots as path components after than an error.

Everything joining two paths has to either use an implementation of that path joining algorithm, but only if they want to joins paths in the potentially dangerous way allowing leading "../". Otherwise they can just use the normal join and even if an attacker gets those dots that will just cause the generated URL to be rejected as invalid.

Of course using a secure implementation is now even more inconvenient thanks to RFC 3986 being around: If you have no control over the generator of relative references, it is always possible that they generate relative references with ".." components after non-dot components.

And if you check all code to properly filter out "/../", keep in mind that convienence does not stop there. After all it is not unheared of for server implementations to helpfully normalize unicode characters, too, or translate them to their nearest ASCII equivalents. Or translate percent escaped characters back before doing path splitting. Or you might think there was some unicode codepoints between those two dots, but they that those were some meaningless control characters that can be omitted. So you need some really restrictive allow lists...

Sync to Apple Health. Share your data like your exercise records, sleep, vitals, steps, and more from Google Health to Apple Health.

The former employees said many Apple files they had been shared on over their careers at the company—including planning documents for product launch events—continued to sync to their personal devices through the iCloud storage service after they left Apple. In some cases, they even received notifications about fresh updates to the documents. Some former employees said they were petrified to delete the files for fear that doing so would attract Apple's attention.

This case is about OpenAI employees wrongfully taking Apple's secret and confidential information regarding our unreleased technologies, processes, and products. Nothing in the filing relates to documents shared by, or stored in, iCloud.

1981 was a different time for computing. It was expensive (both hardware and software), and it was far from a given that machines from one vendor would be able to talk to those from another. In fact, Columbia University had just such a problem, so in 1981, Frank da Cruz and Bill Catchings designed a serial protocol they called Kermit. Because of the many quirks of the DEC-20 and IBM mainframes, the Kermit protocol was highly adaptable from the start: able to handle systems that had trouble processing more than 96 bytes of data at once, able to transfer 8-bit files over 7-bit links, able to translate between character sets (ASCII and EBCDIC then; now also various Unicodes), and of course, handling of error-prone serial links.

Kermit spread rapidly; by 1982, Kermit had been ported to MS-DOS and Unix. Eventually, C-Kermit (an implementation of Kermit in C) became the flagship Kermit. It gained TCP support, an interactive CLI, a powerful scripting language (with features from the shell, Lisp, and expect), and optimizations for today’s high-speed links, such as jumbo packets, sliding windows, and streaming modes. Along the way, Kermit flew on the International Space Station, ran data collection from sensors during hurricanes, and many other uses including postal systems, Boeing 787 manufacturing, and more.

Today, I use it as a powerful ssh wrapper (letting me easily transfer files through multiple nested ssh, sudo, su, etc. commands), a BBS client, to exchange data with me HP 48GX calculator, and so on. It’s also used today to transmit firmware updates to embedded devices. And, of course, anyone that works with vintage systems is likely to use Kermit at some point.

It wouldn’t be until the late 1990s that the TCP/IP stack was finally adopted by most OS vendors, establishing something of a common basis for communication. Of course, we assume this today. Though transferring large files between OSs (say, Linux, Windows, MacOS, Android, iPad, etc.) is still a challenge, even though they all speak TCP/IP! I find that the easiest way to get large files from two computers is to spin up Kermit (see ckwin for a Windows fork of C-Kermit) and just set up a TCP connection over the LAN. In fact, I added a new show interfaces command in C-Kermit 11, making it easy to see your system’s local IPs.

For most of its history, Columbia’s Kermit project was self-funded. Columbia charged for commercial use, which limited its inclusion in Linux distributions. In 2011, 30 years after its founding, Columbia canceled the Kermit Project and released C-Kermit as Open Source under a BSD license. Frank da Cruz, who had still been working with the Kermit project all those years, volunteered to continue maintaining Kermit outside Columbia, and continued development with alpha and beta releases through his retirement from the project in 2025.

I dive into this C codebase

As Debian maintainer of Kermit, I noticed some areas where it wasn’t matching modern expectations. One area was, not surprising for a project of its age, security. Another area was that its character set or line-ending conversions are usually not desired now; we are used to byte-identical binary transfers, and the defaults caused confusion and even some rare instances of data corruption. So I started making a few patches last year.

I’ve worked with old C codebases before, such as Varnish. I’ve generally hated it. You usually find a mix of bad and terrible practices, unclear memory management, and so forth.

But I’ve been living in the C-Kermit codebase for a few months now, and I enjoy it. Yes, this thing is still designed to build on VMS, OS/2, and with compilers that haven’t heard of ANSI — and those that require modern practices. (That em-dash was mine; I knew how to use them before LLMs existed and I’m not going to stop just because LLMs have copied people like me! No AI was used for this post.)

The there is an elegance in all of that. As I worked, I fixed a bunch more potential security issues, both with memory safety and with protecting against a malicious remote in roughly the same manner that some patches to scp did a few years back. I added IPv6 support, of course conditionally compiled because some systems C-Kermit builds on have never heard of IPv6 and never will. (And, of course, with fallback algorithms at runtime for systems that have IPv6 support but not IPv6 connectivity.)

I added unit tests and Python-based end-to-end tests, running nearly 2000 test cases in total. Along the way, I found and fixed a number of bugs going back decades. I learned about FIONREAD being broken on macOS, about NetBSD’s bugs in the pty driver, and fixed bugs in the Kermit protocol implementation itself. I added compatibility tests with the gkermit and ekermit (embedded) implementations, as well as the last full release, C-Kermit 9.0.302 from 2011 (which was difficult to get compiled on a modern system).

There is an extensive changelog describing all the improvements in C-Kermit 11.

C-Kermit development had never really used a VCS at any point, though Kermit veteran Jeffrey Altman imported historical releases into a Git repo, along with some patches that hadn’t made it into a release (which I also pulled in.) There was a lot of disabled code behind #ifdef COMMENT, along with commentary describing why it was no longer used. With Git, we would now generally just remove the old code and explain why in a commit message. I went through and did so with a lot of it, meaning that, at last check, C-Kermit actually has fewer lines of code now than it used to.

Towards a new release

It became apparent pretty quickly that I was making more changes than would make sense as a Debian patch series. Not only that, but they would be more widely applicable to more than just Debian and Ubuntu users. As Linux and BSD distributions were running everything from the last non-beta release (2011’s 9.0.302) to the last beta release (about 1.5 years ago), depending on their different policies about running betas, even sharing patches in a useful fashion was going to be quite difficult.

So, I spun up a project at Open Kermit to coordinate future development in the open and keep Kermit going.

With modern CI, I run that test suite on Linux (x86_64 and arm64), macOS, FreeBSD, NetBSD, and OpenBSD. It builds binary releases on all those platforms, plus a statically-linked Linux binary built with musl libc.

You can download the latest C-Kermit release, and of course contribute to C-Kermit and its website.

Dedication

Frank da Cruz was directly involved with Kermit for 44 years. I’m not aware of any other Open Source project founder being involved for so long. Richard Stallman started working on GNU Emacs in 1984, 3 years after Frank started working on Kermit, but Richard hasn’t been in that role since around 2008.

Accordingly, C-Kermit 11 bears this dedication:

I dedicate this release of C-Kermit to Frank da Cruz.

Frank was directly involved with Kermit for 44 years, from its initial design in 1981 all the way through 2025. He maintained Kermit as an Open Source project after Columbia University ended its sponsorship. I know of no other Open Source project where the founder remains so personally involved for so long.

When Kermit was begun, transfers between different hardware and operating systems were difficult or impossible. Frank helped build a bridge. Kermit glued systems together, from the International Space Station to pocket calculators, and set a new standard for interoperability. It continues to do so.

Kermit is still one of the quietly-working pillars of computing today, enabling everything from firmware upgrades to radios. And, yes, it still reliably transfers files over serial lines.

As we start to spend a lot of time in the Kermit codebase, we do so standing on the shoulders of a giant. Thanks, Frank, for your decades of work on Kermit.

John Goerzen, July 2026

14-Inch MacBook Pro

16-Inch MacBook Pro