diff --git a/libcontainer/logs/logs.go b/libcontainer/logs/logs.go
index 95deb0d6ca7..349a18ed383 100644
--- a/libcontainer/logs/logs.go
+++ b/libcontainer/logs/logs.go
@@ -4,10 +4,19 @@ import (
 	"bufio"
 	"encoding/json"
 	"io"
+	"os"
 
 	"github.com/sirupsen/logrus"
 )
 
+// IsLogrusFd returns whether the provided fd matches the one that logrus is
+// currently outputting to. This should only ever be called by UnsafeCloseFrom
+// from `runc init`.
+func IsLogrusFd(fd uintptr) bool {
+	file, ok := logrus.StandardLogger().Out.(*os.File)
+	return ok && file.Fd() == fd
+}
+
 func ForwardLogs(logPipe io.ReadCloser) chan error {
 	done := make(chan error, 1)
 	s := bufio.NewScanner(logPipe)
diff --git a/libcontainer/setns_init_linux.go b/libcontainer/setns_init_linux.go
index e891773ec57..d1bb12273c0 100644
--- a/libcontainer/setns_init_linux.go
+++ b/libcontainer/setns_init_linux.go
@@ -15,6 +15,7 @@ import (
 	"github.com/opencontainers/runc/libcontainer/keys"
 	"github.com/opencontainers/runc/libcontainer/seccomp"
 	"github.com/opencontainers/runc/libcontainer/system"
+	"github.com/opencontainers/runc/libcontainer/utils"
 )
 
 // linuxSetnsInit performs the container's initialization for running a new process
@@ -117,5 +118,23 @@ func (l *linuxSetnsInit) Init() error {
 		return &os.PathError{Op: "close log pipe", Path: "fd " + strconv.Itoa(l.logFd), Err: err}
 	}
 
+	// Close all file descriptors we are not passing to the container. This is
+	// necessary because the execve target could use internal runc fds as the
+	// execve path, potentially giving access to binary files from the host
+	// (which can then be opened by container processes, leading to container
+	// escapes). Note that because this operation will close any open file
+	// descriptors that are referenced by (*os.File) handles from underneath
+	// the Go runtime, we must not do any file operations after this point
+	// (otherwise the (*os.File) finaliser could close the wrong file). See
+	// CVE-2024-21626 for more information as to why this protection is
+	// necessary.
+	//
+	// This is not needed for runc-dmz, because the extra execve(2) step means
+	// that all O_CLOEXEC file descriptors have already been closed and thus
+	// the second execve(2) from runc-dmz cannot access internal file
+	// descriptors from runc.
+	if err := utils.UnsafeCloseFrom(l.config.PassedFilesCount + 3); err != nil {
+		return err
+	}
 	return system.Exec(name, l.config.Args[0:], os.Environ())
 }
diff --git a/libcontainer/standard_init_linux.go b/libcontainer/standard_init_linux.go
index c09a7bed30e..d1d94352f93 100644
--- a/libcontainer/standard_init_linux.go
+++ b/libcontainer/standard_init_linux.go
@@ -17,6 +17,7 @@ import (
 	"github.com/opencontainers/runc/libcontainer/keys"
 	"github.com/opencontainers/runc/libcontainer/seccomp"
 	"github.com/opencontainers/runc/libcontainer/system"
+	"github.com/opencontainers/runc/libcontainer/utils"
 )
 
 type linuxStandardInit struct {
@@ -258,5 +259,23 @@ func (l *linuxStandardInit) Init() error {
 		return err
 	}
 
+	// Close all file descriptors we are not passing to the container. This is
+	// necessary because the execve target could use internal runc fds as the
+	// execve path, potentially giving access to binary files from the host
+	// (which can then be opened by container processes, leading to container
+	// escapes). Note that because this operation will close any open file
+	// descriptors that are referenced by (*os.File) handles from underneath
+	// the Go runtime, we must not do any file operations after this point
+	// (otherwise the (*os.File) finaliser could close the wrong file). See
+	// CVE-2024-21626 for more information as to why this protection is
+	// necessary.
+	//
+	// This is not needed for runc-dmz, because the extra execve(2) step means
+	// that all O_CLOEXEC file descriptors have already been closed and thus
+	// the second execve(2) from runc-dmz cannot access internal file
+	// descriptors from runc.
+	if err := utils.UnsafeCloseFrom(l.config.PassedFilesCount + 3); err != nil {
+		return err
+	}
 	return system.Exec(name, l.config.Args[0:], os.Environ())
 }
diff --git a/libcontainer/utils/utils_unix.go b/libcontainer/utils/utils_unix.go
index 220d0b43937..842f9b0a6d2 100644
--- a/libcontainer/utils/utils_unix.go
+++ b/libcontainer/utils/utils_unix.go
@@ -7,8 +7,11 @@ import (
 	"fmt"
 	"os"
 	"strconv"
+	_ "unsafe" // for go:linkname
 
 	"golang.org/x/sys/unix"
+
+	"github.com/opencontainers/runc/libcontainer/logs"
 )
 
 // EnsureProcHandle returns whether or not the given file handle is on procfs.
@@ -23,9 +26,11 @@ func EnsureProcHandle(fh *os.File) error {
 	return nil
 }
 
-// CloseExecFrom applies O_CLOEXEC to all file descriptors currently open for
-// the process (except for those below the given fd value).
-func CloseExecFrom(minFd int) error {
+type fdFunc func(fd int)
+
+// fdRangeFrom calls the passed fdFunc for each file descriptor that is open in
+// the current process.
+func fdRangeFrom(minFd int, fn fdFunc) error {
 	fdDir, err := os.Open("/proc/self/fd")
 	if err != nil {
 		return err
@@ -50,15 +55,66 @@ func CloseExecFrom(minFd int) error {
 		if fd < minFd {
 			continue
 		}
-		// Intentionally ignore errors from unix.CloseOnExec -- the cases where
-		// this might fail are basically file descriptors that have already
-		// been closed (including and especially the one that was created when
-		// os.ReadDir did the "opendir" syscall).
-		unix.CloseOnExec(fd)
+		// Ignore the file descriptor we used for readdir, as it will be closed
+		// when we return.
+		if uintptr(fd) == fdDir.Fd() {
+			continue
+		}
+		// Run the closure.
+		fn(fd)
 	}
 	return nil
 }
 
+// CloseExecFrom sets the O_CLOEXEC flag on all file descriptors greater or
+// equal to minFd in the current process.
+func CloseExecFrom(minFd int) error {
+	return fdRangeFrom(minFd, unix.CloseOnExec)
+}
+
+//go:linkname runtime_IsPollDescriptor internal/poll.IsPollDescriptor
+
+// In order to make sure we do not close the internal epoll descriptors the Go
+// runtime uses, we need to ensure that we skip descriptors that match
+// "internal/poll".IsPollDescriptor. Yes, this is a Go runtime internal thing,
+// unfortunately there's no other way to be sure we're only keeping the file
+// descriptors the Go runtime needs. Hopefully nothing blows up doing this...
+func runtime_IsPollDescriptor(fd uintptr) bool //nolint:revive
+
+// UnsafeCloseFrom closes all file descriptors greater or equal to minFd in the
+// current process, except for those critical to Go's runtime (such as the
+// netpoll management descriptors).
+//
+// NOTE: That this function is incredibly dangerous to use in most Go code, as
+// closing file descriptors from underneath *os.File handles can lead to very
+// bad behaviour (the closed file descriptor can be re-used and then any
+// *os.File operations would apply to the wrong file). This function is only
+// intended to be called from the last stage of runc init.
+func UnsafeCloseFrom(minFd int) error {
+	// We must not close some file descriptors.
+	return fdRangeFrom(minFd, func(fd int) {
+		if runtime_IsPollDescriptor(uintptr(fd)) {
+			// These are the Go runtimes internal netpoll file descriptors.
+			// These file descriptors are operated on deep in the Go scheduler,
+			// and closing those files from underneath Go can result in panics.
+			// There is no issue with keeping them because they are not
+			// executable and are not useful to an attacker anyway. Also we
+			// don't have any choice.
+			return
+		}
+		if logs.IsLogrusFd(uintptr(fd)) {
+			// Do not close the logrus output fd. We cannot exec a pipe, and
+			// the contents are quite limited (very little attacker control,
+			// JSON-encoded) making shellcode attacks unlikely.
+			return
+		}
+		// There's nothing we can do about errors from close(2), and the
+		// only likely error to be seen is EBADF which indicates the fd was
+		// already closed (in which case, we got what we wanted).
+		_ = unix.Close(fd)
+	})
+}
+
 // NewSockPair returns a new unix socket pair
 func NewSockPair(name string) (parent *os.File, child *os.File, err error) {
 	fds, err := unix.Socketpair(unix.AF_LOCAL, unix.SOCK_STREAM|unix.SOCK_CLOEXEC, 0)