Liquid Nitrogen
The LN2 attack involves cooling the MBA rapidly down to minus 321 farenheit and allows to preserve the RAM contents for quite some time, potentially allowing the whole-disk-encryption key recovery from RAM, even in a powered off MBA.
The SSD on the MBA, being an SSD, never actually quite deletes data, the trim algorithm presumably keeps writing the whole-disk-encryption key all over the place, such that even a NIST military 'data destruction' overwrite on the SSD isn't actually guaranteed to overwrite your sensitive stuff.
Some Three letter Acronym organisations glue/seal items to block USB & other ports for their staff devices. (a 'cheap' Apple A1305 or similar DVI adapter could have the bare TB connector removed and superglued into the port, I wouldn't do anything more aggressive than that to an Apple MB/MBA)
Machines that are über-protected in any of these ways are easily persuaded to reveal their contents via social engineering = targeted Phish APT or by essential system upgrade components being subverted ( = iTunes upgrades allegedly used by FinFisher in the past) or by generic *.* Certificate Authority SSL certificates (which are still in use for Enterprise and National security means)
The various whole disk encryption schemes might be assumed to have essential third-party maintenance access capabilities anyway.
The sound that your keyboard makes when you type your decrypt passwd can be used to 'guess' it, likewise many keyboards radiate sufficient RF for the keypresses to be scanned from a short distance (I've seen reading at three floors distance in a hotel! - with around $1K of Ettus products used)
It's best to have a vanilla machine , use it normally and store your secrets in a safe! (or saran-wrap covered 64GB microSD card retained in your mouth)
For realistic data security you have to assume your opponents are already in your system, so use multiple independent elements of security.
There are some situations where these 'paranoid' levels of security are necessary - journalism comes to mind in some countries, but on the whole I do trust my national authorities with all of my data.
Ross Anderson has another 600 pages on the subject here
Cambridge UK