(@0xcafebabe: ADHD high-five, I've got almost the same target list, except I'm playing with their NPUs)

Of course it is relevant privacy-wise (Google still gather a lot of metadata about who speaks to who and when), but that's not even the reason I'm mentioning it.

In the RCS specification, there are three lines on a killer-feature: device attestation. A server can whitelist which devices are allowed to connect to it. And of course Google uses this, allowing only Google-certified devices and Apple devices.

And then, there is actually one RCS client for Google's RCS servers. Because Google Messages doesn't use RCS, it uses a custom protocol based on protobuf. My personal guess is that this protobuf is a 1-to-1 matching with actual RCS, and it's both-way compatible. But still, that means that potentially the Google device-attestation won't work with an RCS client.

Sibling comments say that GrapheneOS plan on implementing it. I think it can reasonably work. TBH I'm expecting that giving a LLM the publicly available information about RCS on microG should give a working "send message" within a day (even when using Google apps rather than microG, it's just that the microG work explains the API). And once RCS work in GrapheneOS' message app, it should be pretty straightforward to port to microG, so yay. However I have to admit I'm not optimist about how long it will keep working. I'd say we are two years away from Google enforcing RKP device integrity for RCS, and uh, good luck passing that.

Of course, I'm hoping that, in the EU, the DMA will break this Google/Apple-only device-attestation, but I'm not aware of anyone pushing that ATM.

This has been validated on Qwen3.6-35B-A3B and Qwen3.8-Flash-Next, but with the caveat that it was always on quants.

https://github.com/vagrillo/llama.cpp/blob/moe-expansion/ben...

(Technically it says "which intentionally emits and/or receives radio waves for the purpose of radio communication", I'll let HN crowd determine if Ethernet emits/receive radio waves in an enclosed channel called Ethernet cable)

Two things:

- to rain on the parade, the European RED directive makes secure upgrades mandatory for anything connected to the internet (I suspect that's why Elgato Key Light Mini has signed firmwares). So OEMs are now required to prevent you from doing that. (EN18031-1). It even requires that network credentials (WiFi SSID/PSK) to be stored on secure storage (idk if you can pass that requirement without secure boot. I would guess Elgato does?). "secure upgrade" is loosely defined as "integrity and authenticity are valid at the time of installation" so this requirement doesn't forbid us from upgrading our hardware, but the most likely implementation of OEMs does.

- When you want to do that on Android smartphones (please do!): I recommend to go through GSI/Treble route: This way you quickly have an OS that boots. There are a lot of things to fix, but it will be mostly userspace stuff, which will be easier for the agent to work with. Agent will be able to decompile OEM's userspace and compare with AOSP's userspace, and implement the differences. (That's compared to the ""legacy"" or LineageOS official method which are more convoluted, including kernel stuff, and getting just to "it boots" can be complicated).

(Yes ok, RSA4096 is technically a matter of cost, you just need an infinite amount of money)

I don't know if it's a useful answer to people saying this kind of stuff, but here are some examples of other attacks arbitrary USB pwn allows.

A USB device can appear as a network adapter and most OS will happily route all your traffic there, so your speaker can know which porn you're looking at!

It can also appear as a DisplayLink dongle, so it can see what's on the screen (it does require those specific drivers installed, and uh yeah, no way in hell it's technically possible on that MCU).

It can also turn it into a mouse jiggler to prevent lock screen (yes it's technically the same thing as your first point, just HID, but different angle).

It can also appear as a USB-storage: You don't trust the cloud, so you're writing those super secret documents to give to your boss on the USB drive you just plugged in? Surprise, you actually sent it to the attacker.

[1] Because any AC-powered Zigbee device is a repeater, so just a bulb or a plug is enough

F_gravitational = G m1 m2 /r^2

g = G Mass_earth / r_earth^2

Mass_earth = r_earth^2 * g/G

Density_earth = r_earth^2 * g/G / V_earth

Density_earth = 3*g / (4*Pi*G*r_earth)

Prior to Cavendish we already new g and r_earth, just missing G.

I can't really judge whether 1000 charges is a reasonable target for a car, though i think that 1000 fast charges is reasonable. It should probably be able to push to 5000 slow charges and 500 fast charges, which should fit a lot of use-cases.

I'll add to your comment that it isn't a bug of MCP itself. MCP doesn't specify what the LLM sees. It's a bug of the MCP client.

In my toy chatbot, I implement MCP as pseudo-python for the LLM, dropping typing info, and giving the tool infos as abruptly as possible, just a line - function_name(mandatory arg1 name, mandatory arg2 name): Description

(I don't recommend doing that, it's largely obsolete, my point is simply that you feed the LLM whatever you want, MCP doesn't mandate anything. tbh it doesn't even mandate that it feeds into a LLM, hence the MCP CLIs)