To be fair, the likelyhood of a backdoor is fairly low since individual compute clusters of AI training GPUs are usually always firewalled from the open internet making remote access or even "calling home" very hard.
The actual malicious threat that people overlook (because they're distracted by worrying about backdoors ) is poison pill logic in the silicon. Since the US knows that nearly everyone who buys the H20 GPUS are in China, there could be malicious logic in the silicon that will return bad data or inefficient solutions to data matching certain criteria. For example if they want to hobble China's biotech sector and the parameters feed into Nvidia GPUs resemble DNA sequences, then it could maliciously create a model where it outputs nonsensical results for a new drug compound query.
We've seen countless instances of US sabotaging an adversary's infrastructure in way more clever ways than what's obvious to regular people - Stuxnet-Iran enrichment centrifuge sabotage, Siberian Pipeline explosion of 1982, just to name a couple.
By constantly hyping up a backdoor/remote-access threat diverts attention from more plausible and more malicious threats that actual tech developers/engineers worry about.
Last edited:
