"On a separate note do you think that Amazon refused the request because it turns out they've got recordings of everything from the factory QA testing forward for every device?"
IMHO, probably not.
For that to be the case, Amazon would have to upload the data at some point, and the first thing privacy wonks do with a device like this is throw a up a sniffer and watch all of the traffic these things emanate.
If they were seeing anything like a constant data stream or unreasonably large periodic bulk flows from an Alexa device to the mother ship, they'd have screamed about it.
Given what the device does, the outbound data flows will follow fairly predictable patterns if it's truly behaving as advertised.
"...and then you have two potential problems: first, what does the company giving you the service do with that info and, second, what would third parties (e.g., law enforcement or hackers) like to do with that info if they could get a hold of it. "
Actually, you have at least 3 potential problems: the two above, plus: How long will it be before Amazon is presented with - or is compelled to produce - an "Alexa, Law-Enforcement" version of their software for targeted installation on these devices, along with the new, standard issue Rule-41 based warrant + gag order?
The code to build an Alexa on a Raspberry Pi is already on Github, it's not a stretch to tweak it from "watch for keyword and upload next 30 seconds of audio" to "upload all audio."
Amazon didn't say "come back with a warrant" out of the goodness of their hearts. They don't want US Government to kill what could end up being their flagship product in its infancy.
Ben Wittes has (had?) a blind faith in the inherent "goodness" of the US Government, based on a vastly different set of starting assumptions.
Now, he's being forced to revisit some of his first principles. This is a good thing, because he's respected in his communities in ways that groups like this one are not, which means in theory he has an ability to influences said communities.
Expect some fairly sharp changes in mentality from pundits in the next couple of years. Hopefully they don't come too late to make a difference, although I expect that they have.
But fundamentally, if we want anything resembling a secure IoT, we're going to have to figure out a way to make it more expensive for companies to ship a vulnerable product than it is for them to fix it first, because the attack surface isn't going to get smaller.
here's a more solid start, based on use of MITRE's CVE system.
Assume Samsung is selling IoT enabled toasters, because why not. Everything's better with a network stack. Anyway, MSRP on this toaster is $100usd and Samsung releases the product Jan 1, 2017, and ships 1000 toasters.
Now, if there are no open CVE's on any component of the IoT stack on this toaster in the 90 days before Samsung ships, they're effectively insulated from liability. Oh, and in that world, the sky is Fuscia.
But, If there _is_ an open CVE was announced >= 90 days before Samsung launches the product, _and_ it gets exploited, Samsung is the hook for 5% of the MSRP for each unit sold of said product for every 90 days of age on the CVE.
Example: Samsung begins selling their IoT enabled toaster (MSRP == $100usd) on Jan. 1, 2017. And they sold 1000 of them on day 1. Said toaster has a vulnerability that was announced on Aug. 15, 2016 (just outside the 90 day grace period). If one of these toasters gets exploited and causes trouble, Samsung is going to write a check for (5% of $100) == $5 for each of the 1000 toasters sold as of the date of the CVE being exploited, plus the same fine going forward for each non-patched unit they sell.
Now, pretend that vuln wasn't released on Aug. 1, 2016, it was release on Aug. 1, 2016. Same ship date, same quantity. Except now instead of 5% per toaster, it's 10%. Add 5% for every 90 day interval of CVE age. Also, allow the total penalty per unit to exceed 100% of MSRP with no upper bound. So, you release an IoT enabled toaster with a 12 year old ssh vuln, and it gets exploited? assume qty 4-90 day periods / year to make it easy, now your penalty is (48 * $5) = $240 * 1000 = $240k in fines for each $100MSRP toaster you sold.
And why use MSRP as the basis for the penalty? Well, because it's both easy to validate and publicly verifiable.
No grace period, no appeal, cut a check to a high school to fund a secure coding class, because CVE's are public and theres no way the organization "couldn't have known".
Oh, and multiple CVE's? 5% per CVE, and scale it out.
If you can verifiably patch these toasters 100% then you restart the clock from the time the patch was pushed to the toaster. If you can't patch them, well, eventually you'll get to write a check big enough to make the board pay attention.
Bonus: Specifically disallow said penalties as a loss for tax purposes.
As to your other question: It's a Samsung toaster running a google code, Samsung pays. It's their label. If Samsung wants to go back and fight it out with Google based on contract terms, that's fine, Samsung can attempt to recoup their (already paid) losses from Google.
(yeah, I know. There's no chance this or anything like it will ever happen.)
Kill someone remotely from 25 feet and you can be a long way away before it's even realised that the insulin pump didn't simply malfunction, but was manipulated.
Assuming it can be determined the pump was manipulated. Which isn't a given.
Insulin pumps have two delivery modes:
Bolus, which is used to deliver a large dose of insulin - for example to correct for high blood sugars or to dose for carbs in a meal;
Basal, which is a slow, continuous dosage intended to keep blood sugars level over time. _and_ which, on this model of pump, can be automatically by adjusted based on time of day.
So, all you realistically would need (in theory) would be line of sight, since the 25' limitation is a bluetooth spec limitation and not a hard and fast physical limitation, and to know what time the person typically goes to bed.
I would think a hacker with murderous intent would be much more likely to use a weapon, not a computer.
A weapon is a state of mind, not an object. You can be beaten to death with the (trivially) detachable seatbelt on an airplane if you put your seatmate in a mind to do so.
An insulin pump is no different. It would, however, be damn near impossible to prove or identify after the fact. There's no such thing as "insulin poisoning", there's just "hypoglycemia, resulting in unconsciousness, followed by death" if not caught in time.