If I've understood correctly, the test used was something along the lines of: two non-chinese accounts send back-and-forth messages including benign content (maybe a picture of a pie) and less-benign content (maybe "I like Falun Gong almost as much as I like the CDC"). Then, chinese accounts would observe that the picture of the pie might be censored. I don't actually see an unambiguous description of what messages were sent, or how strong the effect is. Those questions don't really matter for addressing "are non-chinese accounts monitored?" (unless CL is outright lying, obviously yes), but they do matter for the fun question:

Can this be used to construct an attack on WeChat? Providing targeted misinformation for training, to suggest correlations where there really are none, thus triggering WeChat to have a higher false-positive rate when censoring messages?

I'm reminded of RMS's famous practice (and a script in emacs, IIRC): automatically append various keywords to the bottom of emails to screw with any US surveillance that might be getting too nosy.

EDIT: I'm a dope, and should have read the full report (here: https://citizenlab.ca/2020/05/we-chat-they-watch/#part-2---t...) before commenting.

As far as I can tell, an attack isn't possible. WeChat would ideally like to analyze each document or video sent for sensitive content, but it takes some time (on the order of 20 seconds, maybe), and so that analysis can't be performed before messages are supposed to be delivered. However, if WeChat has already seen the video, then it can make the judgement quickly, and perform real-time censorship. Thus, sending a sensitive video the first time won't be blocked, but all subsequent sends will be. CL's result seems to be that if the first video send is between non-chinese accounts, that's still enough to get it analyzed and blocked the second time.