So far the theory is sound, though not unchallenged: outside a few well-studied systems the empirical evidence remains thin, and several researchers dispute its general reach. The problem, however, arises earlier, in the popular translation, where "optimal intermediate level" becomes "tendency towards mildness". These are not the same thing. The optimum can be very high indeed. It depends on a single parameter, and it is the one almost nobody names.
The parameter is temporal, and can be written as a ratio: the time a pathogen needs to kill, divided by the length of the infectious window.
If the ratio is close to one or below it, lethality cuts into transmission. The patient dies, or takes to bed, while still at the height of their capacity to infect. Every more aggressive variant loses infections against its competitors, and selection works against virulence. Here the popular intuition holds.
If the ratio is much greater than one, nothing happens. The patient has already infected everyone they were going to infect long before falling ill. Their death is, from the virus’s point of view, an event without consequence. No selective pressure. Lethality comes free.
Everything else follows from this ratio.

Ebola is the example everyone cites. It spreads through close contact with bodily fluids; viral load rises when the patient is already prostrate and stays very high in the corpse. Someone who is infectious cannot walk, and someone who cannot walk does not meet strangers. It should be said, though, that the evolutionary argument here is weaker than commonly believed: humans are an accidental host, the reservoir is animal, and selection on virulence in humans is nearly absent. Moreover the high infectivity of corpses pushes in the opposite direction to the trade-off.
Above all, "self-limiting" does not mean "burns out quickly on its own". The West African epidemic that began in late 2013 produced 28,652 confirmed, probable and suspected cases and 11,325 deaths across ten countries. For the first time the virus entered densely populated capitals: Freetown, Conakry, Monrovia. The international health emergency was declared in August 2014 and lifted only in March 2016. It took two years and an enormous international effort. It did not stop by itself.

HIV is constantly invoked to show that "clever" viruses do not kill quickly. The example proves the opposite. Untreated HIV kills practically everyone: if the rule were "pathogens that are too lethal die out", it should not exist. It does exist, and it has infected tens of millions of people. Final lethality is not the constraint. Time is.
That said, HIV is also the case in which trade-off theory has found its best empirical support. The trait measured is set-point viral load, the level that stabilises during the asymptomatic phase. It varies between individuals and is partly heritable: those who are infected tend to settle at a value close to that of the person who infected them. A high value makes a person more infectious per act but speeds progression to AIDS; a low value does the reverse. A genuine, measurable trade-off.
Christophe Fraser and colleagues calculated, in a 2007 study in PNAS, that the value maximising overall transmission potential is around 4.52 log10 copies per millilitre. The means observed in real cohorts were 4.36 in the Dutch one and 4.74 in the Zambian one. The virus has settled around the point that suits it. And that point is intermediate, not minimal: it does not converge on harmlessness, it converges on the value that pays best, and that value kills all the same.

There is more. A 2016 study in eLife, based on a long prospective Ugandan cohort, documented genuine attenuation, with set-point viral load falling over twenty years. But the same authors note that in Europe the opposite dynamic has been hypothesised, with adaptation towards a higher value. Same virus, two continents, two directions. If attenuation were a law of nature, this could not happen.
With chronic hepatitis the evolutionary argument falls apart entirely. Cirrhosis and liver cancer from hepatitis B or C arrive after twenty, thirty, forty years of infection. By then the carrier has transmitted the virus for decades and has already had children. That pathology is invisible to natural selection, exactly as diseases that strike after reproductive age are. This is not a virus that has mellowed out of convenience: it is a virus that presents the bill when the bill no longer affects its spread.
Then there are the cases that go openly the other way.
Measles has circulated in our species for at least a thousand years and has not softened by a gram. It has a basic reproduction number between 12 and 18, among the highest known, and remains a major cause of child mortality where vaccination does not reach. What has changed is not the virus. It is us.
Rabies kills those who develop symptoms with a lethality close to a hundred per cent, and persists undisturbed in animal reservoirs as it always has.
Malaria travels on a mosquito, and the mosquito does not care whether the human can walk. Quite the reverse: a feverish, immobile patient is easier to bite. In vector-borne infections selection can reward aggressiveness rather than punish it. The same holds for cholera, which passes through contaminated water and has no need of a patient on their feet.
The textbook case most often cited in support of attenuation, the myxoma virus introduced into Australian rabbits in 1950, deserves to be read in full. Initial lethality approached 99.8 per cent. It did fall, but it settled at intermediate grades that were still extremely high, in some phases it rose again, and in parallel it was the rabbits that evolved resistance. The virus did not become harmless: it stopped where stopping suited it.

The most instructive case, however, does not involve humans, and must be told carefully because it is regularly distorted. Marek’s disease virus in chickens has become far more virulent over the decades. In 2015 Andrew Read’s group at Penn State, with Venugopal Nair of the Pirbright Institute, gave an experimental explanation in PLOS Biology. The vaccines used on farms are "imperfect" or "leaky": they protect the bird from disease but do not prevent infection and transmission. In unvaccinated birds the most lethal strains kill the host before it can spread them, and they die out with it. In vaccinated birds those strains find a host that survives and keeps shedding virus. Selection is turned on its head.
It is the cleanest demonstration that virulence can rise rather than fall. But what that study does not say must be stated at once, because online it is made to say the opposite. The authors themselves write that when a vaccine prevents transmission, as nearly all vaccines used in humans do, this evolution towards greater virulence is blocked. Read’s concern is about partially protective veterinary vaccines, not childhood immunisation. And later work from the same field, in 2020, showed that even a leaky vaccine substantially reduces viral load, both in vaccinated animals and in the unvaccinated ones they infect.

Now apply the ratio to SARS-CoV-2. The infectious window lasts roughly five or six days, concentrated around symptom onset, with a substantial share of transmission occurring before symptoms appear. Death, when it comes, comes two or three weeks after infection. The ratio is of the order of three or four. By the time the patient deteriorates, the virus has already changed host. Lethality does not touch its reproductive success: there is no trade-off to optimise, because there is no cost to pay.
And indeed the data show no trajectory towards mildness. Delta was more severe than Alpha, not less. The virus did not soften by degrees: it jumped back and forth.
Omicron is less severe, that much is true, and it does have a lower tropism for the deep lung. But it did not prevail because it was milder. It prevailed because it evaded immunity and replicated better in the upper airways. Lower severity is a by-product of that tropism, not the cause of its success. Had a more aggressive variant gained the same advantage in early transmission, it would have taken hold just the same.
There is a way to test whether the fall in mortality was due to the virus weakening or to population immunity. You need a place hit by Omicron but without protection. Hong Kong provides exactly that, for tragic reasons.

Until early 2022 the city had contained the virus almost completely: a little over two hundred deaths in two years, and an immunologically naive population. Overall vaccine coverage was high, but not where it mattered: about 42 per cent of those over 65 had two doses, and only 7 per cent of those over 60 had received a booster, barely 2 per cent among the over-80s.
Then Omicron BA.2 arrived. More than nine thousand deaths in three months, one of the highest per capita rates of the entire pandemic.
These are the decisive numbers. Among unvaccinated over-80s, case fatality with Omicron was 21.7 per cent, with a confidence interval between 17.1 and 26.8. In the earlier waves, with the ancestral strain, it had been 24.9 per cent, with an interval between 20.9 and 29.3. Practically the same. The authors of the study, published in Emerging Infectious Diseases, conclude that the intrinsic severity of BA.2 may not be much lower than that of the original strain.
In the same wave, same virus, same weeks: over-80s who had completed the primary series had an estimated case fatality of 11.1 per cent, roughly half. Caution is needed on that figure, though, because vaccinated people in that age band were few and the confidence interval is very wide, from 4.2 to 22.6: taken on its own, that comparison is not enough.
The solid number lies elsewhere. A US CDC report on the same epidemic calculated that the risk of death among unvaccinated over-60s was 21.3 times that of people who had received two or three doses. The variant was the same for everyone. The vaccine was not.

At the root of the misunderstanding lies a compression. A phenomenon with many dimensions is flattened onto one, "dangerousness". At least five dimensions need keeping apart.
Transmissibility. The basic reproduction number is the number of infections generated by one case in a wholly susceptible population: a theoretical number, valid only for a naive population. The one observed in the field also depends on immunity and behaviour. Confusing the two is the most common error in public debate.
Lethality, for which two distinct numbers are needed. The first is the ratio of deaths to infections: the true one, hard to estimate, requiring serological surveys. The second is the ratio of deaths to diagnosed cases: easy to calculate, and systematically inflated when testing is scarce. In March 2020 the second figure in Italy approached ten per cent. The first was below one. It was not the virus that had changed. It was the denominator.
The route of transmission. Aerosol, droplets, faecal-oral, blood-borne, sexual, vector-borne, contact with fluids. This is the dimension that governs all the others, because it determines whether the pathogen needs a host on their feet.
Temporal structure. Incubation, latent period, length of the infectious window, share of pre-symptomatic transmission, time to severe outcome. This is the dimension almost nobody names in public, and it is the decisive one.

Chronicity and reservoir. Acute or chronic, human-only or zoonotic. Smallpox was eradicated because it had no animal reservoir. Influenza never will be, because it lives in waterfowl and pigs.
A note on vocabulary is finally needed, because three words are used as though they were synonyms. Virulence is the damage the pathogen does to the host. Pathogenicity is the capacity to cause disease rather than silent infection. Aggressiveness is not a technical term: in ordinary speech it blends replication speed, tissue damage and lethality, which are three separable things. A virus can replicate very fast and do little harm. A slow one can destroy an organ over thirty years.
No law of nature was protecting us. The virus had no interest in sparing us, because it infected us before it hurt us, and so our fate was none of its concern.
The study by the World Health Organization’s European office, published in The Lancet Respiratory Medicine, estimates that between December 2020 and March 2023 vaccination directly saved more than 1.6 million lives among people over 25 in the European region, with a range between 1.5 and 1.7 million. Recorded deaths were 2.2 million; without vaccines, the model suggests, they would have been around four million. Sixty per cent of the lives saved fall in the Omicron period.

It is an estimate produced by a model, with its assumptions, and the exact number can be debated. It should also be said that a third factor contributed to the fall in mortality, the immunity accumulated from earlier infections, and that Omicron genuinely does have less tropism for the deep lung. None of this is in doubt.
But the direction of causation is clear. And the way to establish it is not intuition. It is to look at what happens to those who were not protected when the virus was the same for everyone.

References
Anderson R.M., May R.M., "Coevolution of hosts and parasites", Parasitology, 85(2):411-426, 1982.
Fraser C., Hollingsworth T.D., Chapman R., de Wolf F., Hanage W.P., "Variation in HIV-1 set-point viral load: epidemiological analysis and an evolutionary hypothesis", PNAS, 104(44):17441-17446, 2007.
Blanquart F., Grabowski M.K., Herbeck J. et al., "A transmission-virulence evolutionary trade-off explains attenuation of HIV-1 in Uganda", eLife, 5:e20492, 2016.
Read A.F., Baigent S.J., Powers C., Kgosana L.B., Blackwell L., Smith L.P., Kennedy D.A., Walkden-Brown S.W., Nair V.K., "Imperfect Vaccination Can Enhance the Transmission of Highly Virulent Pathogens", PLOS Biology, 13(7):e1002198, 27 July 2015.
Boots M., "The Need for Evolutionarily Rational Disease Interventions: Vaccination Can Select for Higher Virulence", PLOS Biology, 13(8):e1002236, 2015.
Bailey R.I. et al., "Pathogen transmission from vaccinated hosts can cause dose-dependent reduction in virulence", PLOS Biology, 18(3):e3000619, 2020.
Meslé M.M. et al., "Impact of COVID-19 vaccination programmes in Europe: lives saved and lessons learned", The Lancet Respiratory Medicine, 2024.
"Epidemiology of Infections with SARS-CoV-2 Omicron BA.2 Variant, Hong Kong, January-March 2022", Emerging Infectious Diseases, 28(9), 2022.
"COVID-19 Mortality and Vaccine Coverage - Hong Kong Special Administrative Region, China, January 6-March 21, 2022", MMWR, 2022.
Istituto Superiore di Sanità, EpiCentro, "Ebola virus epidemic 2014-2016".