We build declaw.ai — sandboxing infrastructure for AI agents, on Firecracker microVMs. We run untrusted code we don't write and can't predict, so when Dirty Frag dropped our first question was: does our isolation boundary hold? We tested it on a deliberately unpatched kernel. It held. Here's why.

The exploit is a page-cache write primitive: it tricks the kernel into overwriting the in-memory contents of any file (/usr/bin/su, /etc/passwd) and gives root. Fully deterministic, no race.

Why it matters for multi-tenant platforms: the page cache is shared across the whole machine. Containers share the host kernel, and namespace isolation, seccomp, and dropped capabilities are all enforced by that kernel. A kernel exploit doesn't need to escape the container — it operates below the layer where container isolation exists. Same structural issue as Dirty COW (2016) and Dirty Pipe (2022). On the day a zero-day drops, before any patch exists, every container-based sandbox sharing that kernel is exposed. Patching closes the window after the fact; it can't close it in advance.

We ran the public PoC (ESP path, CVE-2026-43284) in two environments.

Test 1 — container sandbox (Docker, seccomp on, unprivileged uid=1001, host kernel 6.8.0): unprivileged user to root in under 2 seconds. Seccomp was active but didn't help — the required syscalls were permitted by the profile. With root we read /etc/shadow, host kernel boot params, and Docker overlay2 paths.

Test 2 — Firecracker microVM (unpatched guest kernel, no seccomp, started as root with full capabilities — intentionally MORE permissive than test 1). The exploit worked inside the guest, but every attempt to reach the host failed: host kernel not visible, host processes invisible (the guest has its own kthreadd/kswapd), all host ports closed, only virtual block devices, no host hardware identity. The page cache it corrupted belongs to the guest's own kernel, mapped to a bounded region of host memory via EPT.

The asymmetry is the point: the microVM started with more privilege than the container and still couldn't reach the host. What matters isn't what permissions the software grants — it's whether the kernel is shared. To escape Firecracker you'd need a bug in the VMM (~50K lines of Rust) or KVM; Google's kvmCTF pays $250K for a guest-to-host escape and only one has ever been publicly demonstrated.

If you run untrusted code multi-tenant, the question for any isolation provider: if code inside the sandbox becomes root, can it reach the host or other tenants? If the answer is "as long as we're patched" — that's the gap.

PoC: https://github.com/V4bel/dirtyfrag Full writeup (commands + output): https://declaw.ai/blog/dirty-frag-microvm-isolation