Firmly agree. A lot of sibling comments are talking about environment mutation (which does have issues); I want to talk about environment
read access.
The environment is a standard, locate-able, read-only at runtime k/v store in every process. That makes it an incredibly juicy target for exploits. There are tons of remote exploits well short of RCE which can access all or part of a server process's environment. If that environment contains secrets for everything that process might do, that's asking for trouble.
Consider a user-facing webserver with a rarely-used, admin-only route that talks to AWS APIs. Unless it's deployed on AWS and using IMDS, the 12-factor best practices say there should be AWS credentials in its environment.
Consider a service which, at startup, opens a connection to a telemetry/logging system, then drops privileges and handles requests. 12-factor best practices say there should be a secret for that telemetry system in its environment.
Additional examples abound. Most applications (even ones that aren't internet-facing web servers) use configured secrets infrequently--often only once, to open connections to external services--and not during the vast majority of requests they serve, but we put all secrets in the environment anyway.
Vaults don't automatically solve this problem either; many vaults provide secrets to applications by injecting them into process environment at start.
Good secret management at runtime should ideally be:
1. Mutable or at least delete-able. I really wish there were ways to remove environment variables after they're used (so I could say "once you have an authenticated, open socket or a refreshable auth token to $service, remove the initial login secret from memory entirely"), but absent highly complex multi-process/re-exec dances, that doesn't really exist. If, in Python, you 'del os.environ["foo"]', you haven't modified the environment segment of your program's memory.
2. Not in one common/uniform memory area or key-value API. Hell, it's slightly preferable to have secrets be stored piecemeal in regular variables in memory scattered around your code. Those are going to be slightly harder to find for malware that gets a foothold--security by obscurity, true, but the environment memory block/API is such a tempting and easy target that it buys you a bit more than a false sense of security here.
3. Ideally, stored or encrypted in memory (for secrets that have to stay in memory) such that an exploit which can read process memory doesn't get them for free. Some vaults have a host-local sidecar which provides secrets or a decryption key for them; that way, if an attacker gets memory-read without RCE they can't just exfil a memory image and figure out the decryption key later, but you don't have to be reliant on a remote networked service's uptime for all secret accesses. Even if you don't go that far, securing secrets in-memory at least gives you the option of doing zero-trust stuff based on request payloads, or even just making good-hygiene backend APIs that encode "you can only read the value for secret X if the request is for an admin route and authenticated" (which is a good idea for internet-exposed services with seldom-used risky secrets anyway, but doesn't help with parts 1 and 2 if that API is just wrapping env.get() or whatever).