Bottom line
Inference, medium confidence: adopt a maintained isolation substrate (Firecracker microVMs, gVisor, or Kata) or an agent SDK that delegates isolation to a substrate (E2B on Firecracker microVMs; Cloudflare Sandbox SDK on Cloudflare's own edge substrate), and wrap it with isolation tests of your own. Writing your own isolation layer would recreate years of adversarial test work, but no repository in the evidence publishes a sandbox-escape test suite, and the agent-facing SDKs test their API layer while leaving isolation to the substrate. Before you claim "contained", write the threat model, pin the substrate and guest image by hash, and run a red-then-green confinement suite against that pin.
Adopt, do not rebuild
License checked on 2026-09-23; non-permissive licenses are named in the text.
- firecracker-microvm/firecracker Apache-2.0: the microVM substrate under E2B-class sandboxes, with dedicated security, performance and functional integration suites; rebuilding it means rebuilding its jailer, seccomp and vulnerability-regression tests too. [Verified] (ecosystem/pickup/_evidence/sandbox-exec.md:14
The microVM substrate under E2B/Daytona-class sandboxes
; ecosystem/pickup/_evidence/sandbox-exec.md:14dedicated tests/integration_tests/security/, performance/, functional/ suites
) - google/gvisor Apache-2.0: the application-kernel (syscall interposition) alternative to microVMs, with syscall and runtime suites for isolation-layer correctness; the second isolation paradigm, available instead of writing a third. [Verified] (ecosystem/pickup/_evidence/sandbox-exec.md:15
Application-kernel sandbox (runsc) — the syscall-interposition alternative to microVMs
; ecosystem/pickup/_evidence/sandbox-exec.md:60Conformance-style syscall/runtime suites for isolation-layer correctness
) - kata-containers/kata-containers Apache-2.0: lightweight-VM container runtime with a multi-arch CI matrix and a stress and soak suite kept apart from functional tests, for teams that want VM isolation behind a container interface. [Verified] (ecosystem/pickup/_evidence/sandbox-exec.md:16
Lightweight-VM container runtime (CNCF); multi-arch CI matrix
; ecosystem/pickup/_evidence/sandbox-exec.md:61Stress/soak/stability suite separate from functional tests
) - e2b-dev/E2B Apache-2.0: open-source agent sandbox SDKs on a Firecracker fleet, with one test workflow per SDK surface and timeout and kill lifecycle tests; the ready-made "run this code" API layer. [Verified] (ecosystem/pickup/_evidence/sandbox-exec.md:9
SDKs (js/python), code-interpreter + desktop sandboxes, Firecracker-based microVM fleet
; ecosystem/pickup/_evidence/sandbox-exec.md:55Per-package test workflows (one workflow per SDK/surface, not one mega-workflow)
; ecosystem/pickup/_evidence/sandbox-exec.md:49Timeout + kill lifecycle tests
) - cloudflare/sandbox-sdk Apache-2.0: sandboxed environments on Cloudflare's edge, isolated by Cloudflare's own substrate rather than Firecracker, gVisor or Kata, with a scheduled daily performance workflow and examples for coding agents; an option when your agents already run on Workers. [Verified] (ecosystem/pickup/_evidence/sandbox-exec.md:12
runs sandboxed code environments on Cloudflare's edge; has a scheduled daily performance.yml
; ecosystem/pickup/_evidence/sandbox-exec.md:65they delegate isolation to the substrate (Firecracker microVMs / Workers isolate / Docker)
; ecosystem/pickup/_evidence/sandbox-exec.md:18Cloudflare Sandbox SDK ships examples for claude-code, codex, opencode
)
Copy these practices
- Confinement tests that can fail: assert the jailed process's uid/gid, exact permission bits, device nodes and rlimits; check that the seccomp filter blocks a forbidden-syscall set; diff host and guest CPU vulnerability mitigations, as Firecracker's security suite does. The companion adds a red-then-green requirement: the suite must fail on the unhardened setup and pass after. Starter kit: Acceptance shape: positive observable + planted negative + no-claim line (provisional). [Verified] (ecosystem/pickup/_evidence/sandbox-exec.md:26
Jailer / chroot confinement tests: assert uid/gid, exact file permission bits, device nodes inside the jail, and rlimits
; ecosystem/pickup/_evidence/sandbox-exec.md:27Seccomp filter validation: verify the BPF filter actually blocks the forbidden syscall set
; ecosystem/pickup/_evidence/sandbox-exec.md:28Known-vulnerability regression: run third-party Spectre/Meltdown checker inside the guest and diff host-vs-guest mitigation state
; ecosystem/pickup/pickup-sandbox-exec.md:204A port that cannot demonstrate a red-then-green cycle fails the gate
) - Test against an independent reimplementation: E2B's signed-URL test recomputes the signature with a second crypto library so the test cannot pass by echoing the implementation. It tests auth, not isolation; copy the idea, not the label. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/sandbox-exec.md:31
independent reimplementation of the signature (WebCrypto vs node:crypto) to catch implementation-coupled tests
; ecosystem/pickup/_evidence/sandbox-exec.md:66secure.test.ts is about signed-URL auth, not isolation
) - Performance compared against a baseline, and run on a schedule: Firecracker's PR performance pipeline is configured as A/B comparisons against a baseline; Cloudflare schedules named cold-start, burst and backup-restore scenarios daily. Starter kit: Incumbent oracle pinned before implementation. [Verified] (ecosystem/pickup/_evidence/sandbox-exec.md:39
Buildkite A/B performance pipeline: perf tests run as A/B comparisons against baseline on PRs
; ecosystem/pickup/_evidence/sandbox-exec.md:41Scheduled daily performance workflow: cold-start, sustained-throughput, bursty-traffic
) - Limits tested by behavior, not by config presence: drive the block and network rate limiters and assert real throttling; run code past its timeout and assert the kill. Starter kit: Acceptance shape: positive observable + planted negative + no-claim line (provisional). [Verified] (ecosystem/pickup/_evidence/sandbox-exec.md:47
Rate limiter tests (block/net I/O throttling): drive rate limiter, network rate limiter
; ecosystem/pickup/_evidence/sandbox-exec.md:48Execution timeout test at SDK layer
) - Dangerous end-to-end tests behind an approval gate, local tests without cloud: Cloudflare splits a merge queue from a privileged-PR workflow so risky e2e only runs with extra approval; Vercel ships a mock package for local testing. Starter kit: none. [Verified] (ecosystem/pickup/_evidence/sandbox-exec.md:57
Merge-queue + privileged-PR split (dangerous e2e tests only run with extra approval)
; ecosystem/pickup/_evidence/sandbox-exec.md:58Mock package for local testing without cloud infra
)
Build only if
- Your threat model requires running native code inside one process, where microVM and application-kernel substrates do not fit, as with a from-scratch engine's JIT path. Even then, the capsule packet names V8's in-process sandbox as the deployed reference, which a from-scratch engine cannot borrow, and Wasm runtimes as the way to avoid running native code at all. [Inference] (packets/franken_native_capsule-assessment.md:65
a classic V8-sandbox-shaped problem, scoped here to a from-scratch engine
; packets/franken_native_capsule-assessment.md:151FrankenEngine is not V8 and cannot borrow it
; packets/franken_native_capsule-assessment.md:152is to not run native code at all — run Wasm through a validated compiler
) - A host without KVM does not by itself justify building: the Firecracker battery depends on KVM, and gVisor is the evidenced non-microVM paradigm (the pack does not record whether gVisor needs KVM; check its platform options before relying on it). [Inference] (ecosystem/pickup/pickup-sandbox-exec.md:391
security/performance battery (Buildkite-hosted, KVM-dependent)
; ecosystem/pickup/pickup-sandbox-exec.md:138Two isolation paradigms coexist as norms for the type
) - License gives no reason to build: a fresh read on 2026-09-23 (
gh api repos/<owner>/<repo> --jq .license.spdx_id, plus the LICENSE file for cloudflare/sandbox-sdk, where the API said NOASSERTION) returned Apache-2.0 for all five adopted repositories. [Verified] (stack/METHOD.md:79Check the license of everything you tell a builder to adopt
)
Where FrankenSuite touches this
- franken_native_capsule executes JIT-compiled machine code inside a guard-paged write-xor-execute capsule, and the only code it will run is a 39-byte pre-audited probe; it states it is not a process sandbox or supervisor, its README says the path "establishes no production containment or performance claim", and the repository ships no license text. TRL 3–4, Explore. [Verified] (packets/franken_native_capsule-assessment.md:13
Rust trust boundary for executing JIT-compiled machine code inside a guard-paged W^X (write-xor-execute) sandbox
; packets/franken_native_capsule-assessment.md:13the only machine code it will ever execute is 39 bytes of pre-audited x86-64
; packets/franken_native_capsule-assessment.md:112issuer signature verifier, durable nonce store, process sandbox/supervisor
; packets/franken_native_capsule-assessment.md:143establishes no production containment or performance claim
; packets/franken_native_capsule-assessment.md:13It ships no license text, no CI, no releases
; synthesis/00-overview.md:74franken_native_capsule | 3–4 | Explore | none
) - franken_remote plans seccomp-BPF sandboxing for its Linux software decoder with automated escape-attempt tests [Maintainer claim]; the test file's existence was verified at the pin, not its execution [Verified]. That is the escape-test shape the agent SDKs above do not publish, inside a project with no complete application yet. TRL 3–4, Explore; license: MIT with a rider barring OpenAI, Anthropic and anyone acting for them. (packets/franken_remote-assessment.md:134
the plan demands seccomp-BPF sandboxing for the Linux software decoder with automated escape-attempt tests
; packets/franken_remote-assessment.md:134decoder_sandbox_escape.rs — existence verified at pin
; packets/franken_remote-assessment.md:19There is no complete application, no qualified live transport
; packets/franken_remote-assessment.md:21the license rider names OpenAI and Anthropic as Restricted Parties
; synthesis/00-overview.md:81franken_remote | 3–4 | Explore | Rider
) - franken_engine treats containment as a runtime decision ladder (allow, challenge, sandbox, suspend, terminate, quarantine) with a red-team scenario corpus, inside a JS/TS execution substrate rather than an OS isolation layer. TRL 4–5, Explore; license: MIT with a rider barring OpenAI, Anthropic and anyone acting for them. [Maintainer claim] (packets/franken_engine-assessment.md:119
drive the six-rung ladder allow / challenge / sandbox / suspend / terminate / quarantine
; packets/franken_engine-assessment.md:122(scenario, manifest) pairs pinning expected containment verdicts
; packets/franken_engine-assessment.md:3MIT License (with OpenAI/Anthropic Rider)
; synthesis/00-overview.md:69franken_engine | 4–5 | Explore | Rider
) - frankenlibc is an
LD_PRELOADvalidation membrane over host glibc; its packet reads it as a possible layer for sandboxing agent-generated code, which is analyst inference, not demonstrated use. TRL 4, Explore; license: MIT with a rider barring OpenAI, Anthropic and anyone acting for them. [Inference] (packets/frankenlibc-assessment.md:281The concrete agent workload that would pick this over glibc is sandboxing untrusted AI-generated code
; packets/frankenlibc-assessment.md:3MIT + OpenAI/Anthropic rider (non-OSI; see §4.8)
; synthesis/00-overview.md:91frankenlibc | 4 | Explore | Rider
)
What we cannot say
- Whether any incumbent resists escape: the pack found no sandbox-escape pentest in any tree, and the agent SDKs publish no escape-attempt tests, so every "isolated" claim here rests on substrate regression tests (ecosystem/pickup/_evidence/sandbox-exec.md:65
pentest was found in any repo's tree
; ecosystem/pickup/_evidence/sandbox-exec.md:65None of the agent-facing SDK repos (E2B, Cloudflare, Vercel) publish escape-attempt tests
). - E2B's signed-URL test is auth evidence only (ecosystem/pickup/_evidence/sandbox-exec.md:66
secure.test.ts is about signed-URL auth, not isolation
). - gVisor's secbench measures filter cost, not completeness (ecosystem/pickup/_evidence/sandbox-exec.md:71
benchmarks the cost of seccomp filtering, it does not adversarially test the filter's completeness
). - Firecracker's CI runs on Buildkite, which the pack did not observe running (ecosystem/pickup/_evidence/sandbox-exec.md:70
Firecracker CI is primarily Buildkite
); whether a clean-room team can rerun that battery is open (ecosystem/pickup/pickup-sandbox-exec.md:391security/performance battery (Buildkite-hosted, KVM-dependent)
). - The Vercel mock package comes from a young repository (ecosystem/pickup/_evidence/sandbox-exec.md:68
vercel/sandbox is young (201 stars) but pushed today with real structure
). Daytona's open repository is a tombstone, so its stars are legacy demand, not a live incumbent (ecosystem/pickup/_evidence/sandbox-exec.md:67As of June 2026, Daytona's core development has moved to a private codebase
). - Guest-image provenance is unevidenced (ecosystem/pickup/pickup-sandbox-exec.md:386
Guest-image supply chain: no evidence repo showed signed image
). - Workflow files (Cloudflare
performance.yml,merge-queue.yml,pr-privileged.yml; E2B's per-SDK workflows) were read as files; none was observed running (ecosystem/pickup/_evidence/sandbox-exec.md:3All repos verified 2026-09-23 via api.github.com and raw file fetches
). - Pricing was not in the evidence pack; licenses come from the author's fresh read, not the pack. The FrankenSuite rows are packet reads; none of those tests was executed for this verdict (packets/franken_native_capsule-assessment.md:145
the honest evidence base is the test suite (10 tests, read but not executed here)
).
Revisit when
- A maintained, version-pinned public sandbox-escape vector suite appears, or an agent SDK (E2B, Cloudflare, Vercel) publishes escape tests at its own boundary.
- Daytona's core development returns to a public repository, or
vercel/sandboxshows sustained maintenance. - franken_native_capsule gains a consumer and its planned mutation harness, or franken_remote's decoder escape test is observed running.
Sources
- Pickup companion: ecosystem/pickup/pickup-sandbox-exec.md, the charter seed, oracle candidates, and first claim registry for this area.
- Evidence pack: ecosystem/pickup/_evidence/sandbox-exec.md, the repositories checked against the GitHub API, with dates.
- Verdict source: stack/sandbox-exec.md, written under stack/METHOD.md.
Starting a build here?
Begin from the pickup companion for Sandboxed code execution: it carries the charter seed, the oracle candidates and how to verify them, and the first claims to register, so a new project starts from evidence instead of from scratch.