Skip to content

Security: maceip/AXP

SECURITY.md

Security model

Use GitHub's private vulnerability reporting for security issues. Include a minimal reproducer and affected version. Do not publish private session data or credentials in an issue.

The host controls identities, access, event ordering, leases, spending records and review. Contributors can be unreliable or dishonest. Agent output is accepted only from the current lease holder, with a matching epoch, before the lease expires. Contributors cannot approve their own tool requests, change other contributors' budgets or record verification results. Sharing a blob between sessions does not grant access to it.

Git worktrees are not sandboxes. Native mode allows agent tools to run as the local user. ACP capability negotiation does not confine a process's own filesystem or network access. The Docker launcher mounts the worktree, drops capabilities, disables networking and provides a read-only root with temporary storage. To use a remote model from a container, configure a proxy or network access. Networking is disabled by default. TEE attestation is not implemented yet.

AXP control credentials are not inherited by the agent. Provider credentials stay on the machine running the agent. To enforce a hard spending cap, use provider limits or a quota proxy. Usage reports are self-reported. When usage is unknown, the full reservation is charged. A signature identifies who made a claim about an artifact; it does not prove the claim is true.

Use TLS remotely, separate contributor and verifier identities, and isolated verification infrastructure. One live hub owns one database; do not share it between concurrent hosts or expose its files to an agent. The current release keeps all session history, so you are responsible for managing storage.

There aren't any published security advisories