Please report suspected vulnerabilities privately via GitHub's Report a vulnerability button (repository → Security → Advisories), not a public issue. We aim to acknowledge within a few days.
aplan reads a repository's plan; aplan verify resolves evidence against git, and
aplan verify --execute runs named tests. The security-relevant invariants:
aplan checkis hermetic. It opens no network sockets and spawns no subprocess, it reads files and computes. This is enforced by a test (aplan-e2e/tests/hermetic.rs): the crates the check path lives in name no process, socket, or git API, andunsafe_codeis forbidden workspace-wide, so there is no raw-syscall escape hatch. Thejsonschemadependency is pinneddefault-features = falseto disable remote schema resolution.- A task file can never run a command.
aplan verify --executeruns a test, and a task file can arrive in a fork pull request, so the task file supplies a test name, never a command. The name is substituted, as a whole argument, into a runner template defined only in the trusted top-level.aplan.yml, and the command is spawned with an argv vector and no shell. A runner template that would splice the name inside a larger argument (the one way a shell-using runner could be turned against a task file) is rejected when the config loads. So the worst a hostile name can do is fail to match a test. Pinned bya_hostile_test_name_cannot_inject_a_commandinaplan-evidence; a Kani proof of the same property is planned.--executeis off by default and CI must not enable it for fork pull requests. urlevidence is not fetched by default. When enabled it will be HTTPS-only with redirects disabled (SSRF hardening), a timeout, a body-size cap, and a content-addressed cache.
A report that breaks any of these, a way to make check reach the network, or a task
file that causes command execution, is a genuine vulnerability, and the most valuable
kind to send.
Pre-1.0: only the latest commit on main is supported.