diff --git a/docs/internals.md b/docs/internals.md
index b88f9b76..a023e230 100644
--- a/docs/internals.md
+++ b/docs/internals.md
@@ -981,7 +981,10 @@ How it works:
 2. The interpreter is loaded as `ET_DYN` at `g->interp_base` (computed
    dynamically: 60 GiB for 36-bit IPA, 1020 GiB for 40-bit IPA).
 3. `build_linux_stack()` passes `AT_BASE` (interpreter load address) and
-   `AT_EXECFN` (`argv[0]`) in the auxiliary vector.
+   `AT_EXECFN` (the execve filename, supplied by the caller) in the auxiliary
+   vector. Linux takes `AT_EXECFN` from `bprm->filename`, so it stays the
+   program the guest asked for even when `argv[0]` is an alternate name or the
+   rosetta binfmt_misc argv prepends the translator.
 4. The entry point becomes `interp_entry + load_base`; the dynamic linker
    takes over from there.
 5. Guest absolute paths reach the host through `path_translate_at()`
diff --git a/src/core/bootstrap.c b/src/core/bootstrap.c
index 6e79c161..985d1a99 100644
--- a/src/core/bootstrap.c
+++ b/src/core/bootstrap.c
@@ -627,7 +627,7 @@ int guest_bootstrap_prepare(guest_t *g,
     boot->stack_pointer = build_linux_stack(
         g, g->stack_top, stack_argc, stack_argv, (const char **) environ,
         stack_elf, stack_elf_load_base, stack_interp_base, native_vdso,
-        rosetta_execfd, &auxv);
+        rosetta_execfd, elf_guest_path, &auxv);
     if (boot->stack_pointer == 0) {
         log_error("failed to build initial stack");
         free(rosetta_argv);
@@ -881,9 +881,10 @@ int guest_bootstrap_rosetta_post_reset(guest_t *g,
 
     uint64_t native_vdso = vdso_build(g);
     linux_stack_auxv_t auxv;
-    uint64_t sp = build_linux_stack(g, g->stack_top, rosetta_argc, rosetta_argv,
-                                    (const char **) environ, &rr.rosetta_info,
-                                    0, 0, native_vdso, rosetta_execfd, &auxv);
+    uint64_t sp =
+        build_linux_stack(g, g->stack_top, rosetta_argc, rosetta_argv,
+                          (const char **) environ, &rr.rosetta_info, 0, 0,
+                          native_vdso, rosetta_execfd, elf_guest_path, &auxv);
     free(rosetta_argv);
     if (sp == 0) {
         log_error("build_linux_stack failed during exec re-bootstrap");
diff --git a/src/core/stack.c b/src/core/stack.c
index 4a93c5b6..61642d7f 100644
--- a/src/core/stack.c
+++ b/src/core/stack.c
@@ -140,6 +140,7 @@ uint64_t build_linux_stack(guest_t *g,
                            uint64_t interp_base,
                            uint64_t vdso_base,
                            int execfd,
+                           const char *execfn,
                            linux_stack_auxv_t *auxv_out)
 {
     /* Linux initial stack layout (growing from high to low):
@@ -218,6 +219,31 @@ uint64_t build_linux_stack(guest_t *g,
     uint64_t platform_ptr = str_ptr;
     str_err |= write_str(g, platform_ptr, "aarch64");
 
+    /* AT_EXECFN: the filename handed to execve, copied onto the stack as its
+     * own string exactly as fs/binfmt_elf.c does.
+     *
+     * The kernel takes this from bprm->filename, not from argv[0], and the two
+     * diverge in two ways elfuse reproduces: execve(path, "altname", ...)
+     * reports path, and under binfmt_misc the interpreter rosetta.c prepends to
+     * argv is not the program the guest asked to run. Taking the string from
+     * the caller rather than from an argv index keeps that contract out of the
+     * argv layout, which differs between the native and rosetta forms and is
+     * free to change again (see the preserving-form note in rosetta.c).
+     *
+     * Guests that identify themselves through auxv rather than argv[0] depend
+     * on getting this right: rust-coreutils dispatches its multi-call applet
+     * from AT_EXECFN, so a leaked interpreter path makes every applet abort
+     * with "unknown program 'rosetta'".
+     */
+    uint64_t execfn_ptr = 0;
+    if (execfn) {
+        size_t execfn_len = strlen(execfn) + 1;
+        if (!stack_take(&str_ptr, stack_floor, execfn_len))
+            return 0;
+        execfn_ptr = str_ptr;
+        str_err |= write_str(g, execfn_ptr, execfn);
+    }
+
     /* Dynamically allocate pointer arrays to avoid stack buffer overflow with
      * large argument or environment lists. calloc(0, ...) is
      * implementation-defined, so always allocate at least one slot. The extra
@@ -251,8 +277,11 @@ uint64_t build_linux_stack(guest_t *g,
         str_err |= write_str(g, str_ptr, argv[i]);
     }
 
-    /* AT_EXECFN: pointer to argv[0] string (write it near the top) */
-    uint64_t execfn_ptr = (argc > 0) ? arg_ptrs[0] : 0;
+    /* Callers with no filename to report keep the historical argv[0] spelling
+     * rather than an AT_EXECFN of 0, which no Linux process ever sees.
+     */
+    if (!execfn_ptr && argc > 0)
+        execfn_ptr = arg_ptrs[0];
 
     /* Phase 2: Build the structured part of the stack. Align str_ptr down to 16
      * bytes first.
diff --git a/src/core/stack.h b/src/core/stack.h
index e7a60767..f2e1cf9d 100644
--- a/src/core/stack.h
+++ b/src/core/stack.h
@@ -54,6 +54,11 @@ typedef struct {
  * ET_EXEC). interp_base is the load base of the dynamic linker (0 if statically
  * linked). vdso_base is the guest address of the vDSO ELF image (0 if no vDSO).
  * execfd is the pre-opened binary fd for binfmt_misc (AT_EXECFD); -1 if none.
+ * execfn is the guest path handed to execve, reported as AT_EXECFN. Linux takes
+ * that string from the execve filename rather than from argv[0], and the two
+ * differ whenever the caller passed an alternate argv[0] or the kernel
+ * prepended a binfmt_misc interpreter, so callers pass it explicitly instead of
+ * leaving it to be inferred from argv's shape. NULL falls back to argv[0].
  * If auxv_out is non-NULL, it receives the exact auxv words written to guest
  * memory, in the same order exposed by /proc/self/auxv.
  * Returns the initial SP (stack pointer) to pass to the guest.
@@ -68,4 +73,5 @@ uint64_t build_linux_stack(guest_t *g,
                            uint64_t interp_base,
                            uint64_t vdso_base,
                            int execfd,
+                           const char *execfn,
                            linux_stack_auxv_t *auxv_out);
diff --git a/src/syscall/exec.c b/src/syscall/exec.c
index 1d8b343b..8bced5da 100644
--- a/src/syscall/exec.c
+++ b/src/syscall/exec.c
@@ -1545,9 +1545,14 @@ int64_t sys_execve(hv_vcpu_t vcpu,
         uint64_t exec_vdso = vdso_build(g);
         exec_republish_shim_globals_or_die(vcpu, g, verbose);
 
+        /* AT_EXECFN gets the same guest-visible spelling published to
+         * /proc/self/exe below, so the two surfaces agree on what this process
+         * is. Passing path rather than argv_const[0] also matches Linux for
+         * execve(path, ["altname"], ...), where the kernel reports path.
+         */
         sp = build_linux_stack(g, g->stack_top, argc, argv_const, envp_const,
                                &elf_info, elf_load_base, interp_base, exec_vdso,
-                               -1 /* no AT_EXECFD */, &auxv);
+                               -1 /* no AT_EXECFD */, path, &auxv);
 
         /* 0 is build_linux_stack's failure return. Past the point of no return
          * there is no image to go back to, and programming SP_EL0 from it would