I had been working on it for a couple years, using it myself. I finally released it recently. It heavily uses an embedded Lua VM to power its scripting engine.
I had been working on it for a couple years, using it myself. I finally released it recently. A number of users are contributing feature requests, testing, etc on the discord as well.
That's a personal choice. You alone should be able to decide if that's worth the complications it introduces to your taxes and your relationship with your employer. However, the cleanest and most dignified route is to pay for expenses incurred in the course of business. And employees should have the right to refuse to put up their own money for such expenses, without consequence.
I hope they bring back the former UI that allowed you to explicitly toggle "AI / Agent" mode on/off in a terminal session, and gets rid of the Oz / Cloud Agent stuff.
I don't want this auto-detect agent request. The explicit toggle was perfect.
It had a simple UI with a clear button / key combo to toggle the “agent mode”. That, plus the fact it could “warpify” my SSH connections made this a useful utility.
Then a couple months ago they completely changed the UI. It doesn’t work as it once did. My saved prompt templates didn’t work as they did before, the agent toggle was gone (you can now start some ‘/agent’ command but it is much less intuitive) and they seem to be focusing on these cloud agents and code editing.
I want none of these things. I loved a simple terminal that let me still execute sudo, let me ssh into remote machines but still bring Claude and OpenAI models to interact with my session.
argh, sorry that we made it worse for you, that definitely was not the intention. we were trying to make the modes clearer.
You can still basically get back to the same way of working if you set new terminal sessions to be "agent sessions" and enable natural language detection. that's how i use it.
The bigger problem for me is that the realtime voice modes lack tool use, so they can't look anything up or do anything. Model strength definitely also matters, but even dumb models can be helpful when they can look things up and try things out. And smart models that don't do those things kinda suck.
It reminds me of the standard counter to the Chinese Room thought experiment: the person inside doesn’t understand Chinese, but the system _does_. The person, the rules, and the lookup tables together form the thing doing the understanding.
Presence of one or more: unexpected outbound traffic observed via Ethernet, increased battery consumption, interactive response glitching, display anomalies ... and their absence after hard reset key sequence to evict non-persistent malware. Then log review.
What are examples of logs that you're considering IOCs? The picture you are painting is basically that most everyone is already compromised most of the time, which is ... hard to swallow.
By minimizing apps on device, blocking all traffic to Apple 17.x, using Charles Proxy (and NetGuard on Android) to allowlist IP/port for the remaining apps at the router level, and then manually inspecting all other network activity from the device. Also the disappearance of said traffic after hard-reset.
Sometimes there were anomalies in app logs (iOS Settings - Analytics) or sysdiagnose logs. Sadly iOS 26 started deleting logs that have been used in the past to look for IOCs.
How did you determine that a connection was malicious? Modern apps are noisy with all of the telemetry and ad traffic, and that includes a fair amount of background activity. If all you’re seeing are connections to AWS, GCP, etc. it’s highly unlikely that it’s a compromise.
Similarly, when you talk about it going away after a reset that seems more like normal app activity stopping until you restart the app.
That doesn’t have any details supporting the belief that this traffic was malicious or a sign of compromise. I’d easily believe that it’s picking up developer telemetry or ad networks but without some hard evidence this sounds like misinterpretation rather than a compromise.
Traffic was monitored on a physical ethernet cable via USB ethernet adapter to iOS device.
Charles Proxy was only used to time-associate manual application launch with attempts to reach destination hostnames and ports, to allowlist those on the separate physical router. If there was an open question about an app being a potential source of unexpected packets, the app was offloaded (data stayed on device, but app cannot be started).
MDM was not used to redirect DNS, only toggling features off in Apple Configurator.
Surely you used several USB Ethernet adapters to rule them out as being the source as well right? Those types of dongles are well known for calling home.
Good observation :) Multiple ethernet adapters: Apple original (ancient USB2 10/100), Tier 1 PC OEM, plus a few random ones. Some USB adapters emit more RF than others.
It excluded the published hostnames for services and CDNs (some of which resolved to GCP, Akamai, etc) published by Apple for sysadmins of enterprise networks, https://news.ycombinator.com/item?id=46994394. It's indeed possible that one of the unknown destination IPs could have been an undocumented Apple service, but some (e.g. OVH) seem unlikely.
So how did you identify this as a breach? I'm struggling to find this credible, and you've yet to provide specifics.
Right now it comes across as "just enough knowledge to be dangerous"-levels, meaning: you've seen things, don't understand those things, and draw an unfounded conclusion.
Feel free to provide specifics, like log entry lines, that show this breach.
Please feel free to ignore this sub-thread. I'm merely happy that Apple finally shipped an iPad that would last (for me! no claims about anyone else!) more than a few weeks without falling over.
To learn iOS forensics, try Corellium iPhone emulated VMs that are available to security researchers, the open-source QEMU emulation of iPhone 11 [1] where iOS behavior can be observed directly, paid training [2] on iOS forensics, or enter keywords from that course outline into web search/LLM for a crash course.
I worked at Corellium tracking sophisticated threats. Nothing you’ve posted is indicative of a compromise. If you’re convinced I’d be happy to go through your IOCs and try to explain them to you.
Thanks. In this thread, I was trying to share a positive story about the recent iPad Pro _NOT_ exhibiting the many issues I observed over 5 years and multiple generations of iPhones and iPad Pros. If any new issues surface, I'll archive immutable logs for others to review.
With the link I provided, a hacker can use iOS emulated in QEMU for:
• Restore / Boot
• Software rendering
• Kernel and userspace debugging
• Pairing with the host
• Serial / SSH access
• Multitouch
• Network
• Install and run any arbitrary IPA
Unlike a locked-down physical Apple device. It's a good starting point.
I'm much more convinced that you're competent in the field of forensics. But I still don't think suspicious network traffic can be categorically defined as a 'device breach.'
For all you know, the traffic you've observed and deem malicious could just as well have been destined for Apple servers.
Apple traffic goes to 17.0.0.0/8 + CDNs aliased to .apple.com, which my egress router blocks except for Apple-documented endpoints for notifications and software update, https://support.apple.com/en-us/101555
They said upthread that they had blocked 17.0.0.0/8 ("Apple"), but maybe there are teams inside Apple that are somehow operating services outside of Apple's /8 in the name of Velocity? I kind of doubt it, though, because they don't seem like the kind of company that would allow for that kind of cowboying.
I don't doubt it in the slightest. Every corporate surveillance firm—I mean, third-party CDN in existence ostensibly operates in the name of 'velocity'.
There’s no hard evidence that you’ve put forward that you’ve been breached.
Not understanding every bit of traffic from your device with hundreds of services and dozens of apps running is not evidence of a breach.
Have you found unsigned/unauthorized software? Have you traced traffic to a known malware collection endpoint? Have you recovered artifacts from malware?
Strong claims require strong evidence imo and this isn’t it.
As mentioned elsewhere in this thread, traffic from each iOS app was traced via Charles Proxy, the endpoints allowlisted for normal behavior, and finally the app was offloaded so it could not generate any traffic from the device. Over time, this provided a baseline of known outbound traffic from the device, e.g. after provisioning a new device with a small number of trusted apps.
I agree with other posters that you seem to be capable of network level forensics, but you have said nothing to back up what you consider a device breach other than 'some cloud destined network traffic which disapears after a hard reset'.
In my experience of forensic reports, this link is tenuous at best and would not be considered evidence or even suspected breach based on that alone.
I think its the end-to-end, integrated nature of it.
API driven, have "elastic resources", etc, etc. Rather than bolting together various solutions you get to have a "Cloud-like" stack in your own datacenter.
So while most of the software is open source rather than proprietary, you still have a fair point that customers pay for support (as they do with most enterprise products). One could theoretically use the product without first-party software updates, managing the open source oneself... but that would have practical impediments (and runs counter to the all-in-one simplicity that customers value in the Oxide product).
Two points about your last point. First, software improvements benefit all customers; as the business grows, the effective cost per customer shrinks. Also, most customers grow their Oxide deployment or will replace hardware after a depreciation cycle. The sustainability of investments into the software (and the product generally) is on solid ground.
Back in the 90s and 00s, lots of companies churned out software products that were sold once, supported forever. It was a sort of Ponzi scheme, supporting old customers with money from new customers. Which was okay during a period of high growth. But sooner or later the market matures, growth plateaues, and the cost of ongoing maintenance becomes a much bigger problem.
Right now you're growing fast and swimming in VC money, so this is probably not an issue. At some point, though, you might find that even hardware depreciation cycles don't provide as much of a cushion as you hope they will. In an economic downturn, people might suddenly realize that Oxide hardware actually remains serviceable much longer than they expected. :)
https://runemud.com
I had been working on it for a couple years, using it myself. I finally released it recently. It heavily uses an embedded Lua VM to power its scripting engine.