aitm is pre-1.0 and is published to npm as @developerz.ai/aitm. Only the latest published
version is supported. Check yours with aitm --version, and upgrade before reporting against an
older one.
Report privately. Do not open a public issue, pull request or discussion.
Use Report a vulnerability on this repository. It opens a private advisory that only the maintainers can see until a fix ships.
Please include:
- the output of
aitm --version, the runtime and its version (Node or Deno), and your OS; - the command you ran and the relevant part of your config, with keys and tokens removed;
- steps to reproduce, ideally against a throwaway repository;
- the impact: what an attacker gains, and what they have to control first (the target repository's contents, a pull request or review comment, an issue, a config file, the network).
- Acknowledgement within 3 business days.
- A triage decision and a severity within 7 days of the acknowledgement.
- Once it is triaged, we tell you the fix timeline, and we credit you in the advisory unless you ask us not to.
aitm runs on your machine with your credentials: your model provider's API key and your gh
login. It reads the target repository, runs a model's tool calls in a git worktree, opens pull
requests and can merge them.
In scope, for example:
- a credential (your API key, a
ghtoken) reaching a pull request, a commit, a comment or anything elseaitmpublishes, or reaching anyone other than the provider you configured; - a file or search tool reading or writing outside the worktree it was given;
- content someone else controls (a pull request, a review comment, an issue, a file in the target
repository) making
aitmitself merge, push or open something the user did not ask for; aitmmerging past branch protection without--admin, or--admindoing more than the README says it does;- the same classes in
@developerz.ai/ai-claude-compat, which provides the tool surface.
Out of scope:
- what
--adminis documented to do: it bypasses the review requirement, on purpose, for a user with admin rights on the repository; - prompt injection that steers the model's own tool calls inside the worktree. That is a known limit
of every coding agent, and it is a vulnerability here only when it gets past a boundary
aitmenforces: the worktree confinement of the file tools, the merge rules, or credential handling; - a model writing wrong or insecure code into a pull request. Review and CI are there to catch that;
- vulnerabilities in OpenRouter, a model provider, the
ghCLI or the AI SDK. Please report those upstream; - anything that needs prior control of your machine or your accounts.