Skip to content

Security: adeoluo/agent-sandbox

Security

SECURITY.md

Security Policy

Reporting a vulnerability

Please report vulnerabilities privately through GitHub (Security, then "Report a vulnerability") or by email to the maintainer at adeoluo@hotmail.com. Do not open public issues for sandbox escapes.

A useful report includes the policy file, the request (argv or file operation), the OS and kernel version, the output of agent-sandbox inspect --policy <file>, and what the request achieved that the policy should have prevented.

Scope

In scope:

  • A request the engine allows that the policy, as documented, should deny.
  • A path, symlink, or Unicode form that escapes the filesystem rules.
  • A way for a command running under an active OS backend to write outside the writable roots, or (macOS) use the network, or (Linux, Landlock ABI 4+) open a TCP connection.
  • Audit log modifications that audit verify does not detect, other than tail truncation without --expect-head and full-chain rewrites (both documented).
  • Resource limits that do not apply as documented.

Out of scope, because they are documented limitations: anything an allowed interpreter can do that the per-OS table in the README lists as not enforced; behavior with no OS backend; setsid escapes from the process group; running as root.

Controls in this repository

  • SAST: CodeQL for Rust (.github/workflows/codeql.yml).
  • Dependencies: versions are pinned in Cargo.lock; updates are reviewed and applied by hand.
  • Lints: cargo clippy --all-targets -D warnings on Linux and macOS.
  • Unsafe code: limited to libc and syscall wrappers in limits.rs, audit.rs, sandbox.rs, and isolation/linux.rs, each with a SAFETY comment.

There aren't any published security advisories