You raise some excellent points on the categories of existential risks:

Manhattan project

detonating the oceans was an expression of the level of sophistication of scientific knowledge — the scientists at that time knew that we lived right next to a star, and cosmic rays were discovered in 1912 — which even now in 2019 are “terrifyingly and mysteriously energetic” so there is an excellent emperical case to be made that the claim that the earth is not made of compounds that are a “nuclear powder keg” waiting to detonate

LHC

— one in 50 million seems frankly a little too risky for my blood but I think the quoted figure has to be caveated by two things — we know even better than Manhattan Project timed how energetic the universe is, and it’s still here. Further, estimating your likelihood of being metaphyiscally mistaken is a fraught exercise.
Example — 1 in 50 million is more likely than rolling 6 on 10 of 10 dice.
Would I hook a bomb in my house to the outcome of rolling 10 dice? Not if I could help it — but I do comprehend that risk with great certainty.
Should I feel the same over CERN? Only the deluded could claim that. I don’t work at CERN, and nor do I work on this area of risk assessment.
I know that it is yes, creating conditions that are likely very rare in our corner of the universe. Unique in the universe though? — Our galaxy literally orbits a multi million solar mass black hole! ;)

Avian flu?

Messing with that seems crazy to me — we know that infecting cells is literally the primary function of that virus and it’s very good at it. How one could argue that it’s predictable and within safe bounds, given the system is also known to be the most complex kind of system we know of.
Yet, in order to understand it and also understand the risk level we have to do the studies. We are assuming ignorance is worse, which I argue emperically has proven to be case throughout history.

The insurance company model of assessing damage for catatrophes is clearly invalid as it implicitly assumes there will be survivors who will be able to settle the tally. To my mind — on that basis, it’s fatuous to use that model as events that render the concept of continuity moot can’t be assessed using a framework that expresses costs in pre-catastrophe units.

We need a better accounting — I don’t know what, perhaps a moral argument is the only type framework that can be used?