I think that based on the need for constant overwhelming honest transactions and a complete dependence on the proper functioning of an honest network, that an attack is plausible, feasible and probably even likely when the costs of the double-spend attack will be rewarding enough.
With the Tangle, the majority hashpower must be always honest and always working. As researchers have already outlined, "The number of assiduous honest transactions must ALWAYS be found to be in the majority. In order to guarantee that, they must, by all practical estimations, be constantly generating new honest transactions AT ALL TIMES."
The IOTA Foundation already acknowledges this in their FAQ documentation:
The most critical factor ... is the greater adoption of the IOTA technology increasing
the throughput of transactions on the network to meet the fundamental
security assumption - that ... honest transactions is large compared to that which an attacker could feasibly produce.
Low Cost Devices and Ternary Chips Help Attackers Too
One problem is that IoT devices might be many (billions even!) but they are low-power devices and cannot be constantly churning out new "honest" transactions. And that assumes too, that the IoT devices are running to spec and also not hacked with malware and/or susceptible to control by malicious actors.
Another problem is that the same technology that is said to allow IoT devices to provide low-cost PoW per transaction on limited electrical and hardware resources (JINN Trinary Chips) is exactly the same technology that a hacker or attacker can use to overpower the network with fake transactions.
What TPS is actually required to thwart an attacker?
The basis of this is an answer that shows that using current Bitcoin hashrates and resources as a benchmark, the IOTA network would need to generate 1.7 billion transactions per second of honest transactions - constantly and without letup.
This rate of honest transactions required is just impossible because even the network at the most optimized state in the next 100 years probably will not allow that many transactions to circulate in a peer-to-peer network per second.
It is also impossible because the amount of 1.7B TPS needed to be produced will ONLY resist against 1% of the current BTC network hashrate in the power of an adversary.
I would like to emphasize that - 1.7B TPS is required to thwart just 1% of the Bitcoin hashrate in the hands of an adversary.
What about Omnipresence?
It is often stated that the concept of "omnipresence" is required for an adversary to efficiently and successfully distribute their dishonest transactions. This is a false argument because it is proven that with 8 neighbors, the entire network can be efficiently permeated with a single double-spend in only 7 gossip generations (sometimes even 6).
What about Tethering Protections?
IOTA nodes also use mutual tethering rather than automatic peer discovery. This means that nodes cannot discover the rest of the network as easily as Bitcoin and other common true peer-to-peer Gossip-based networks.
However, this fact also lends power to the attacker because the honest nodes cannot see each other and have enough network intelligence to see the attacker's initial attack and ensure that honest transactions are properly distributed at all times.
Although it seems that the network topology is nearly impossible to know, an attacker can create many fake peers and then mutually tether them into the network, thereby gaining a more complete network graph because unlike regular users, the attacker is invested in the attack versus regular individually tethered users that are only using the network and any conveniences it may provide.