Re: [RFC/discuss] memfd_secret(): opt-in visibility for security monitoring (eBPF/audit)

On Wed, Jul 8, 2026 at 8:09 AM BoxStrikesTeam <[email protected]> wrote:
>
> Hi,
>
> While experimenting with mseal() and memfd_secret() together on a
> recent kernel, I ran into what looks like a gap worth discussing:
> content placed in memfd_secret()-backed memory is invisible not just
> to other processes, but also to in-kernel tracing/observability
> mechanisms such as eBPF's bpf_probe_read_user() (EFAULT, as expected,
> since the region is removed from the direct map and kernel page
> tables entirely).
>
> That's working as designed for the threat model memfd_secret targets
> (protect secrets even from a compromised kernel / ROP-based
> exfiltration). But it does mean a local, unprivileged process can
> now allocate memory that:
>
>  1. Is invisible to kernel-side introspection (eBPF tracing,
>  /proc/<pid>/mem, ptrace-based tooling) once populated via
>  memfd_secret(), and
>  2. Can additionally be sealed against mprotect()/munmap()/mremap()
>  via mseal() so its protection bits can't be relaxed either.
>
> Neither primitive alone is new or alarming - mseal() protects mapping
> metadata, not confidentiality, and memfd_secret() is explicitly
> opt-in and disabled by default. But combined, they let a process
> carve out a region that a host-based EDR relying on eBPF
> tracing/probe_read helpers cannot inspect, and cannot force back into
> an inspectable state either. For userland security tooling that
> assumes 'anything a tracing eBPF program can attach to, it can read',
> this is a small blind spot.
>
> I want to be clear I'm not suggesting memfd_secret()'s core guarantee
> should be weakened - the ability to keep data hidden from a
> compromised kernel is the entire point, and forcibly exposing content
> to tracing programs would undermine that model and create a new
> attack surface via the eBPF verifier/helper path itself.
>
> What I'd like to raise for discussion instead is something closer to
> what was already anticipated in the original series - Mike, your
> commit message for memfd_secret() mentions: 'Once there will be a use
> case that will require exposing secretmem to the kernel it will be an
> opt-in request in the system call flags.' Two lighter-weight
> directions that stay consistent with that opt-in philosophy:
>
>  a) An LSM hook / audit event at memfd_secret() call time (this
>  echoes Christian Gottsche's 2022 RFC to label secretmem inodes
>  via inode_init_security_anon for SELinux). Even without content
>  visibility, logging 'process X created a N-byte kernel-invisible
>  mapping' gives EDR/audit systems a behavioral signal to alert on,
>  without touching the confidentiality guarantee at all.
>
>  b) A distinct opt-in flag (as foreshadowed in the original commit)
>  that a process could set to allow a CAP_BPF/CAP_SYS_ADMIN-gated
>  tracing context to read the region for legitimate
>  monitoring/debugging - fully opt-in, off by default, and never
>  implicitly available to unprivileged tracers.
>
> Is (a) something that's been considered further since the 2022
> secretmem-inode-labeling thread? And is there any existing tracking
> issue for the kind of opt-in-visibility idea in (b), or was it
> dropped as not worth the complexity?

Modern kernels do allow LSM labeling of secretmem inodes, see
secretmem_file_create().

-- 
paul-moore.com