Brevi is in public preview and still 0.x. The latest minor release gets every fix. The previous minor gets fixes for Critical and High security issues only, for 4 weeks after its successor ships. Older releases get nothing. See docs/SUPPORT.md.
- No external security audit yet. We have not had an external audit; we are seeking funding for one. Until then, do not run hostile code under the sandbox.
- The sandbox is a policy layer, not an isolation boundary.
--sandbox(the default in thebrevi mcptools and the SDKs, except theirverify, which runs CPython) checks at build time what the program may touch, and, without--allow-pyor--allow native, the program reaches files and the network only through Brevi's own code, which checks the caller's grants. There is no namespace, container or VM around the program, and it runs as the same OS user asbrevi. Only a program that loads CPython gets kernel holds on files and the network as well, on Linux only: Landlock for files (5.13 or later, with Landlock on; where it is missing, named folders are refused and otherwise CPython keeps full file access) and, on x86-64 and arm64, a seccomp filter against the network when none is granted. Arust fnor crate under--allow nativegets no kernel hold at all. The sandbox stops agent code from touching what the caller did not grant; it is not built to hold code that is trying to break out, and a bug in Brevi's checks or runtime can be a way out. - Run untrusted code in a container or VM. Run
breviitself inside one, with no credentials in it, and use the sandbox as a second layer inside it. --allow-pyand--allow nativeare full trust. Once CPython or native code is loaded, the program can do anything the user runningbrevican, apart from the kernel holds and limits the threat model lists. Grant them only to code you would run without a sandbox.
The details, and what each layer does and does not cover, are in docs/THREAT_MODEL.md. How we test it is in docs/SECURITY_TESTING.md.
| Use | What it needs first |
|---|---|
| A design partner running its own trusted agent code inside its existing container or VM | Nothing more: this is a fair use today, with the four points above understood. No audit is needed, because nobody relies on the sandbox as the boundary. |
| A public beta | An independent security review of a fixed scope (the sandbox checks, the MCP server, DLL loading on Windows) is strongly recommended. |
| Offering the sandbox as a boundary for hostile code (not before an audit) | A paid external audit. We will not make that claim before one is done. |
The external audit is on the roadmap; it does not hold up the pilot use above.
Report it privately. Never open a public issue for a vulnerability.
- On the repository's Security tab, choose Report a vulnerability (GitHub private reporting), or
- email support@brevi-lang.org with "security" in the subject.
Include what you ran, what happened, and the Brevi version (brevi --version).
| Step | Target |
|---|---|
| First reply | within 3 days |
| Severity and a fix plan | within 7 days |
| Fix, Critical | 7 days |
| Fix, High | 30 days |
| Fix, Medium | 90 days |
| Fix, Low | best effort |
Severity is Critical, High, Medium or Low. We use CVSS v4 as a guide, not a formula.
Critical and High fixes ship as a patch release as soon as they are ready, even outside the usual Monday release. Medium and Low fixes ship in the next weekly release.
Disclosure is coordinated. We publish the advisory when the fixed release ships, and at most 90 days after your report, unless you agree to a different date. We request CVEs through GitHub Security Advisories (GitHub is a CNA). We credit reporters by GitHub handle unless they ask us not to.
Brevi compiles programs into native binaries, so the most important cases are:
- a program declared with
allowlines, or run withbrevi run --sandboxor through thebrevi mcptools (sandboxed by default), that reads or writes files, reaches the network or the environment without being granted it; - an MCP call that widens its own access: turns the sandbox off or grants itself more than the person who started
brevi mcpallowed (without--trust-calls); - a
--limit(MCPlimits) that does not hold where the threat model says it does; - generated Rust or runtime code that is memory-unsafe for a program Brevi accepted;
- a request that gets past the token or origin checks of
brevi mcp --http; install.shor a release download fetching or running something it should not, or a release artifact that does not match its checksum.
A wrong answer that is not a sandbox escape is a normal bug: please open an issue (CONTRIBUTING.md). What the sandbox does and does not protect against is in docs/THREAT_MODEL.md. Known limits that are documented there are not vulnerabilities.
Every release has a SHA256SUMS file. install.sh checks the archive against it and installs nothing if the check fails. To check by hand:
sha256sum -c SHA256SUMS --ignore-missing
From v0.28.0, each release also carries:
- a CycloneDX SBOM,
brevi-sbom.cdx.json; - GitHub build-provenance attestations for every archive. Verify one with
gh attestation verify FILE -R ets-public/brevi; - a reproducible-build check: the Linux x86-64 binary is rebuilt in a second folder and must be byte-identical.
Earlier releases have only SHA256SUMS.
This covers a vulnerability being exploited, or a compromised release artifact, registry package, workflow or credential. The steps:
- Contain. Revoke the credential. Delete the affected release assets. Yank the version on PyPI and crates.io, deprecate it on npm, and delete the ghcr.io image tag.
- Fix and ship a patch release.
- Publish an advisory that says which versions and artifacts were affected and what users should do.
- Write a short post-mortem in the advisory.
The target is containment within 24 hours of confirmation.
See Where Brevi stands at the top: no external audit yet, the sandbox is a policy layer and not an isolation boundary, and code you have reason to distrust belongs in a container or VM with no credentials. See docs/THREAT_MODEL.md.