> I'd argue that those examples are a higher bar to hurdle than failing to recognize a spear phishing attack, and can be mitigated by solutions like always-on VPN.
The theory behind TLS MITM is that it's an extraordinary and dangerous method that could be justified if sufficiently effective. If there are a dozen common ways to route around it, the risk is more than the benefit.
A VPN can't fix it because a compromised endpoint would be able to choose which traffic it sends over the VPN. Whereas if the endpoint isn't fully compromised then you could be doing whatever scanning is being done by the middlebox on the endpoint itself, without centralizing on single point of compromise for the entire network.
> And if the malware doesn't work because it has certificate pinning, well, that's a win too.
It may not make an outside connection, but that doesn't mean it doesn't work. It could still infect every machine on your internal network. What's the chance that none of them are ever in range of a public WiFi?
This is before even getting to the issue of steganography. Information theory says that if your legitimate communications contain zero entropy then they can be encoded into zero bits, i.e. you don't need a network connection at all, but if they contain nonzero bits of entropy then an attacker can encode that much arbitrary data into the stream and still be indistinguishable from legitimate data.
So the whole thing is inherently a cat and mouse game. A lazy attacker may use a data pattern that isn't found in the legitimate data and then a middlebox vendor may find it and use that to distinguish their traffic, but as soon as they do the attacker can stop using it. The longer the game is played, the better the attackers get at making their data indistinguishable and the fewer remaining undiscovered ways to distinguish it that it's possible to find. In the limit there are none left, the data is completely identical to legitimate data with the same amount of entropy, the attacker is only assigning different meaning to it at the endpoints.
