astorlm is pre-1.0. Security fixes land on main and ship in the next
release; older minor versions are not backported.
| Version | Supported |
|---|---|
0.x (latest) |
✅ |
older 0.x |
❌ |
Please do not open a public issue for security reports.
Use GitHub's private reporting instead: Report a vulnerability.
Include, as far as you can:
- the affected version, entrypoint and platform,
- a minimal reproduction,
- the impact you believe it has.
You can expect an acknowledgement within a few days. Once a fix is ready, it is released and the advisory is published with credit to you, unless you prefer to stay anonymous.
astorlm is a library that lets a language model run tools on your behalf.
Understanding the boundaries matters more than any single report:
- The default
Executorruns commands on the host.LocalExecutorgives the model the same privileges as the process embedding the library. It is a convenience backend, not a sandbox. If the model's input is untrusted, useDockerExecutor(OS-level isolation) or a customExecutorof your own. resolveSafeconstrains paths tocwd, not the shell. The built-inread/write/edittools refuse to escape the working directory, but thebashfamily executes whatever command it is given. Path confinement is not command confinement.QuickJsCodeRunneris the sandboxed tier. It runs self-contained source inside a memory-safe WASM runtime with no filesystem, network or host syscalls unless you grant them explicitly.- Prompt injection is a real risk for any agent. Content the model reads —
files, web pages, MCP tool results — can contain instructions. The
beforeToolExecutionhook is the intended enforcement point: use it to require authorization for sensitive tools. SeerestrictToolsHookfor a ready-made allowlist. - MCP servers are trusted code.
mountMcpServerruns the transport you configure and does not validate tool inputs — the server is expected to. Only mount servers you trust. - Credentials. API keys are read from the environment through
AuthStorageand passed to providers. The library does not persist them. Session files written byFileSessionManagercontain conversation content, which may include whatever the model or your tools put there — treat session directories as sensitive.
Reports about these documented behaviours are not vulnerabilities in themselves, but reports showing a way to bypass a boundary the library claims to enforce definitely are.