F FourIA staging 2026.10.29 GitHub ↗

F4E1 chunk-size bound (decrypt memory safety)

On this page

Status: hardening for a decrypt-time OOM found by adversarial fuzzing. Applies to every runtime that restores an F4E1-encrypted archive — the legacy Sandbox path served by develop and the Sandbox 1.0 (v1-sandbox) track.

The issue

src/crypto/backup-encryption.ts decrypts chunk by chunk:

header (17B) | AES-256-GCM chunk | tag | AES-256-GCM chunk | tag | …

decryptStream buffers a full chunkSize + TAG_LENGTH before it can authenticate and emit that chunk, and parseHeader used to accept any chunk size up to 256 MiB. A corrupt or maliciously written object could therefore make a restore try to buffer ~256 MiB in a single ArrayBuffer — far above the Workers isolate memory limit — and the Worker is OOM-killed mid-restore, exactly when a recovery archive must be readable. The legacy restore path reaches this code on every encrypted recovery point:

  • src/crypto/r2-backup.ts#decryptBackupArchive → chunkSizeFromHeader + decryptStream on backups/<id>/data.sqsh.enc.
  • src/persistence.ts uses decryptBytes for config versions (one chunk).

Fuzz finding: a 17-byte header advertising a 256 MiB chunk is enough to trigger it — no valid ciphertext body is needed.

The change

  • New MAX_CHUNK_SIZE (16 MiB) is enforced by parseHeader, so chunkSizeFromHeader / decryptStream reject an out-of-range header from the header bytes alone, before any chunk is buffered. 16 MiB is four times the 4 MiB writer default and far below the isolate limit.
  • encryptBytes clamps its single-chunk header to MAX_CHUNK_SIZE, so the single-shot writer (config versions, secrets) can never emit a header a reader would reject.

Existing objects are unaffected: the writer default (4 MiB) and every encryptBytes payload small enough to be one chunk are inside the bound. An object written with a header above 16 MiB is already unreadable in practice (that is the OOM case) and now fails fast instead.

Verification

  • src/crypto/backup-encryption.test.ts — default and maximum accepted; 0 and MAX_CHUNK_SIZE + 1 rejected; the historic 256 MiB header rejected from the header alone; encryptBytes clamps and still round-trips a MAX_CHUNK_SIZE + 1 payload.
  • Seeded fuzz harness (bun run --cwd apps/fouria test:fuzz:sandbox-v1 [iterations] [seed]): round-trip, length invariants, single-byte corruption, wrong key, mid-chunk truncation, and the header bound across thousands of random shapes.

Known limitation (unchanged)

F4E1 authenticates each chunk but does not commit to a total plaintext length, so truncating a multi-chunk ciphertext exactly on a chunk boundary decrypts to a shorter plaintext without an authentication error. On the legacy path the squashfs restore and its size/integrity validation are the only guard; the Sandbox 1.0 DirectoryBackup path additionally checks record.size before the SDK restore. A future format version should authenticate a total length/EOF marker.