F FourIA staging 2026.10.29 GitHub ↗

Sandbox v1 / F4E1 adversarial & fuzz report

On this page

Status: adversarial coverage added; four hardening changes landed on v1-sandbox, one of which also applies to the legacy runtime and must ship on develop. This note covers the attacks run against SandboxV1 (apps/fouria/src/sandbox-v1.ts) and the shared at-rest codec (apps/fouria/src/crypto/backup-encryption.ts), the defects found, and which findings apply to the legacy Sandbox path.

How to run

# Codec / lifecycle adversarial suite (deterministic, part of `bun run test`)
bun run --cwd apps/fouria test src/sandbox-v1-codec-adversarial

# The broader DO adversarial suite (recovery point, boot refusal, process map)
bun run --cwd apps/fouria test src/sandbox-v1-adversarial

# Seeded fuzz harness over the F4E1 codec (many more shapes)
bun run --cwd apps/fouria test:fuzz:sandbox-v1            # 500 iterations, fixed seed
bun run --cwd apps/fouria test:fuzz:sandbox-v1 5000 987654

The harness exits non-zero only when an invariant breaks; documented weaknesses are printed as WARN with the seed so a finding is reproducible.

Findings and applicability

#SeverityDefectApplies toFix
1critical (data loss)startContainer() destroyed the container on any setup failure, including a container that was already running (DO restart against a live container, or a retry). A transient intercept()/setInactivityTimeout() error destroyed tenant filesystem state.SandboxV1 only (legacy ensureGateway kills the gateway process, never the live container)Destroy only a container this call started (isNewContainer).
2high (corruption)ensureRunning() guarded on #setup === null || !container.running, and container.running is not guaranteed to flip inside start(). A burst of first requests could call startContainer() several times: duplicate start() and a second intercept(), which the SDK documents as replacing a running backup operation’s grant.SandboxV1 only (legacy boot is owned by start-openclaw.sh’s container lock + the Worker’s per-key start promise)Single in-flight attempt (#starting); concurrent callers join it, and only the initiator clears #setup on failure.
3high (data loss)Concurrent restores of one record shared the transient plaintext key: one restore could delete it while the other’s SDK restore was still queued, failing an intact backup with BACKUP_NOT_FOUND.SandboxV1 only (legacy restores are serialized by the shared R2 backup-lock lease)Per-record restore queue (withRestoreSlot).
4high (OOM)An F4E1 header accepted a chunk size up to 256 MiB. Decryption buffers chunkSize + tag before authenticating it, so a corrupt or maliciously written object could make a restore buffer ~256 MiB and OOM the Worker mid-recovery — far above the isolate memory limit.Shared: SandboxV1 and legacy (r2-backup.ts decryptBackupArchive decrypts .sqsh.enc with the same decryptStream/chunkSizeFromHeader)MAX_CHUNK_SIZE (16 MiB) enforced on read; encryptBytes clamps its single-chunk size so it can never emit a header a reader rejects.

Regression evidence: with sandbox-v1.ts at its pre-fix logic, the lifecycle / restore-race tests fail; with the fixes they pass. The crypto bound has explicit unit tests (default accepted, MAX_CHUNK_SIZE accepted, 0 and MAX_CHUNK_SIZE + 1 rejected, a 256 MiB header rejected from the 17 header bytes alone).

Conclusion for develop (legacy)

Only finding 4 exists in the runtime develop serves. develop carries the same src/crypto/backup-encryption.ts, and the legacy restore path (src/crypto/r2-backup.ts#decryptBackupArchive) calls chunkSizeFromHeader + decryptStream on every encrypted recovery point. A corrupted/forged data.sqsh.enc header can therefore OOM the Worker exactly when a tenant is being recovered. Findings 1–3 are v1-only code paths and stay on this branch.

Documented weaknesses (not changed)

  1. Chunk-boundary truncation is length-preserving. AES-GCM authenticates each chunk but F4E1 has no total-length commitment, so truncating a multi-chunk ciphertext exactly on a chunk boundary decrypts to a shorter plaintext without error. SandboxV1 is protected by withDecryptedDirectoryBackup’s decryptedLength(...) === record.size check before the SDK restore (covered by a test); legacy .sqsh.enc has no recorded plaintext length, so a truncated archive reaches sandbox.restoreBackup() and the squashfs/size validation there is the only guard. A future format version should authenticate a total length/EOF marker.
  2. A container replaced underneath a live DO is not re-instrumented. If the platform restarts the container while the DO incarnation survives, container.running stays true, so ensureRunning() short-circuits and the new container keeps the intercepts registered by the previous instance (the SDK drops per-instance outbound intercepts). The DO cannot observe a container generation through the 1.0 surface; the cutover should add an identity probe.
  3. Off-workerd uploads buffer the object. putKnownLengthStream falls back to collectStream when FixedLengthStream is absent (Node tests only). Production (workerd) takes the streamed path, which the suite asserts.
  4. The idle policy trusts WORKER_URL. Fuzz over arbitrary hostnames, dev-mode values, and durations never produced a non-positive/NaN/over-6h timeout, and billable production always resolves the keep-alive maximum; a malformed hostname degrades to keep-alive (the safe direction).