`&'s T` → `&'static T` in safe code (UAF)

[rucsesec]

I tried this code:

fn static_str_identity()
  -> for<'s> fn(&'s str) -> (
        [&'static &'s (); 0],
        &'static str,
    )
{
    |s: &str| (
        [],
        s,
    )
}

fn main()
{
    let f = static_str_identity();
    let local = String::from("123");
    let s: &'static str = f(&local).1; // <- should be rejected!
    drop(local);
    let _unrelated = String::from("UAF!");
    dbg!(s); // <- compiles and prints `"UAF!"`
}

I expected to see this happen: argument requires that s is borrowed for 'static

Instead, this happened: code compiles successfully and the excution potentially triggers Use-After-Free on the String local.

rogram returned: 0
Program stderr
[/app/example.rs:21:5] s = "UAF!"

This seems very much like a Use-After-Free issue, and it's quite likely related to the compiler.It can be reproduced in the latest version of the Rust compiler.

• Demo: godbolt permalink
• Demo: on current playground

rustc --version --verbose:

rustc 1.84.0-nightly (c1db4dc24 2024-10-25)

note: This issue appears to be different from #114936, which was related to the Fn* Trait and was fixed in version 1.78; whereas this problem can still trigger a vulnerability in the latest version.

[compiler-errors]

This is just another instance of #25860.

Subtyping when storing the return value of static_str_identity() causes us to "cast away" implied bounds of the output type of the function. This allows us to turn the signature from ([&'static &'s (); 0], &'static str) to ([&'first &'s (); 0], &'second str), for two other arbitrary existential regions 'first and 'second, with only the requirement that 'static: 'first and 'static: 'second.

Then later, when we check that the signature is well-formed at the call-site (f(&local)), we only require that 's: 'first due to the well-formedness of the [&'first &'s (); 0] argument, but don't do anything for the 'second region. That allows us to later constrain 'second: 'static without actually "linking" it to 's, which is the lifetime of the local string we're passing in.

[apiraino]

For additional context, T-types is aware of this issue and the current status is summarized at this comment.

@rustbot label -I-prioritize

[lcnr]

Closing in favor of #25860 which we're slowly working towards. Thank you for finding and reporting this issue 👍