I mean like
@Crimson Longinus stated, there are a lot of monster with debugs and damage that gets brutal in large numbers or if played smart.
Well, again, fair enough, and, give me a sec, but, I'll address this in a moment.
As an example, a PC in one of my games got bitten by a death dog, failed his save (despite being a tough barbarian) and was poisoned/diseased for a very long time, which made him extremely vulnerable to later encounters. Death dogs are a CR 1 encounter, there were like five PCs at the time who IIRC were level 2, so it was (by Kobold Fight Club standards, at least) a Trivial encounter. But if he hadn't been lucky enough to shake the poison off the next day, he very likely would have died.
Yes, and this also works with what I'm going to say, so, I'm including it here for clarity.
CR is a predictive model. And, as such, it will never, ever be 100% accurate. It cannot be. There's no way for it to be. Heck, I'm rather impressed it's as accurate as it is to be honest. Imagine the HUGE variation between two groups of characters of the same level. It's enormous. I remember back in the 3e days people complaining about CR back then too but not taking into account the assumptions of CR. CR has so many assumptions built in - standard array PC's, no feats, no magic items, based on the classes that existed in 2014 and nothing that has come later, based on the spells that existed in 2014 and nothing later, so on and so forth.
Does anyone seriously think that two groups of 5 PC's, one built only using the SRD and nothing else and the other built using every single 5e WOTC book are going to have the same power levels? Seriously?
Take the Death Dog example. THat barbarian had to fail a DC 12 Con Saving throw multiple times over the course of several long rests. Presumably we're looking at about rolling an 8 or higher to make the saving throw. And that's presuming no one actually tried to help this character. Every time he fails, he loses a d10 HP off his max HP. We're talking a 2nd level Barbarian here, so, we're talking 23 hp (assuming a 14 Con, not an unreasonable assumption I think). It would take at least 3 days and more likely closer to a week for this to kill this character. The CR system cannot possibly predict this. No predictive system could. The number of failed checks, the complete lack of any help from any other characters, and the DM rolling very high would all have to come together for this character to potentially die from this disease.
How in the world would you, as a game designer, possibly predict any of that? Instead, you go with baselines. It's really, REALLY unlikely that this effect is going to kill a PC. So, it probably doesn't factor at all into the CR calculation of this PC. Why would it?
Now shadows? Sure, anything that bypasses the HP system is automatically going to be more difficult to calculate. That was
@pemerton's point about Tucker's Kobolds. They are 100% a mechanical exploit because everything they do bypasses the standard combat rules and goes diving off into other resolution rules. Which, of course, CR cannot account for.
I mean, sure, kobolds could rig up collapsing roof traps that instantly kill the party. 4d10 damage? Yup, that'll kill low level PC's and seriously hurt others. And it's not like that's unreasonable for them to do. But, again, that's the point of CR. CR doesn't take that into account. Killing PC's isn't all that hard. There's so many things you could do. Zombies infected with Yellow Mold and released at the party - dead PC's. But, again, this is a bit all self-evident. If you dramatically increase the effectiveness of the monsters somehow, yes, they are more dangerous. That's not exactly news.
I mean, good grief, simply using arrow slits would massively increase the lethality of an encounter. Yeah, all baddies have +5 AC and advantage on area saves. Yeah, that'll dramatically up the lethality of the encounter. But, again, the CR system does not, in any way, account for that. The further you deviate from the baseline assumptions of the creature, the less accurate CR will be as a predictive tool.