https://github.com/ErdemOzgen/copy-fail-cve-2026-31431/releases

Contents

• Quick Test (No Build)
• Build Prerequisites
• Local Build
• Usage
• Troubleshooting
• Overview
• Why It Matters
• Root Cause
• Trigger Path in authencesn
• How the Bug Became Reachable
• High Level Exploit Flow
• Fix Summary
• Remediation
• Disclosure Timeline
• How It Was Found
• Repository Layout
• GitHub Actions
• Tested Distributions
• References

Build Prerequisites

• Go 1.23+
• GNU Make

Local Build

Build all Linux targets:

Build and generate checksums:

Output binaries in dist/:

• copy-fail-cve-2026-31431_linux_amd64
• copy-fail-cve-2026-31431_linux_arm64
• copy-fail-cve-2026-31431_linux_386
• copy-fail-cve-2026-31431_linux_armv7
• SHA256SUMS

Usage

Default run:

Use a custom target path:

Disable post run su execution:

Print build metadata:

Troubleshooting

If execution fails at offset 0 with an error similar to:

the kernel most likely rejected the vulnerable AEAD in-place path (for example because the host is patched for CVE-2026-31431 or algif_aead is mitigated/disabled).

Quick checks:

On remediated systems, this PoC is expected to fail.

Overview

According to the public disclosure, the vulnerable path allows an unprivileged local user to influence a 4 byte write primitive into page cache content of readable files under specific conditions.

The important operational property is that the in memory page cache copy can diverge from on disk bytes during runtime, which can make impact analysis not trivial for teams relying only on disk level integrity checks.

Why It Matters

Copy Fail is notable because it was reported as:

• Deterministic (no race requirement in the described trigger path).
• Portable across major Linux distributions in testing.
• Compact in implementation, with no external payload builder dependency.
• Relevant beyond single process scope due to shared page cache behavior.

Root Cause

At a high level, the issue sits at the intersection of three mechanics:

1. AF_ALG AEAD request handling.
2. splice() page cache backed data movement.
3. authencesn decryption scratch behavior.

In the vulnerable design, page cache backed segments can be linked into a request shape that later receives writes during algorithm processing, violating a key assumption that only intended destination regions are modified.

Trigger Path in authencesn

In the documented path, authencesn uses destination side scratch operations around ESN layout handling and performs a write at an offset beyond the expected decrypt output contract.

When combined with the in place AEAD setup and chained scatterlist references, this write can land in page cache backed memory for attacker selected regions.

How the Bug Became Reachable

The disclosure describes a multi commit evolution:

• authencesn behavior originated in earlier integration history.
• AF_ALG AEAD support introduced userspace reachability with splice() interaction.
• Later in place optimization in algif_aead combined source and destination flow in a way that made this intersection exploitable.

The key lesson is compositional risk: individually reasonable changes can become unsafe in combination.

High Level Exploit Flow

The public write up and PoC describe this sequence:

1. Open and configure an AF_ALG AEAD socket for authencesn(hmac(sha256),cbc(aes)).
2. Supply crafted metadata and chunked data.
3. Use splice() to route page cache backed file data through the request path.
4. Trigger decrypt handling and repeat in controlled 4 byte steps.

This repository's Go implementation mirrors that public PoC logic for research and validation in authorized environments.

Fix Summary

The reported fix reverts vulnerable in place AEAD operation in algif_aead to out of place behavior.

Conceptually:

• Before fix: source and destination could collapse into one writable chain for this path.
• After fix: source and destination remain separate, preventing page cache backed source segments from being used as writable destination regions in this flow.

Remediation

• Patch kernels to versions containing the upstream fix.
• Roll out vendor kernel package updates across fleet and images.
• Consider temporary hardening controls (for example, policy restrictions on AF_ALG usage where operationally feasible).
• Validate remediation with runtime checks, not only disk checksum workflows.

Disclosure Timeline

Based on the public write up:

How It Was Found

The published research credits human guided, AI assisted subsystem analysis focused on userspace reachable crypto code paths and scatterlist/page provenance under splice().

Repository Layout

• main.go: Go PoC port.
• Makefile: deterministic multi arch Linux build targets.
• .github/workflows/build.yml: CI build and release artifact workflow.

GitHub Actions

Workflow file: .github/workflows/build.yml

Triggers:

• Pull requests
• Push to main
• Tag push matching v*
• Manual workflow_dispatch

Behavior:

1. Builds all Linux binaries with checksums via make clean checksums.
2. Uploads CI artifacts named with ref + commit SHA.
3. Publishes release assets automatically for tags matching v*.

Suggested release flow:

Tested Distributions

References

• Original public article: https://xint.io/blog/copy-fail-linux-distributions
• CVE entry: CVE-2026-31431