Pullwright runs AI agents that read your code, execute commands on runners and open pull requests, so we treat security reports as a priority.
Report privately through GitHub security advisories. Do not open a public issue or pull request for a vulnerability.
Include what an attacker can do, the steps to reproduce, and the version or commit. You can expect an acknowledgement within three working days.
- Escaping the runner's isolation, or reaching another job's data or tokens.
- Getting the control plane to run code from a repository it plans for. Planning is meant to be read-only.
- Bypassing approval: getting an issue opened or implemented without approval from a linked user with write access.
- Forging Slack, Teams or web requests, session cookies or account-link tokens.
- Leaking secrets (GitHub tokens, the Anthropic key, the app's private key) into logs, issues, PRs or prompts.
- Runners are throwaway containers: non-root, with CPU, memory and PID limits, no-new-privileges, no host mounts, and removed on exit. They receive a GitHub installation token scoped to one repository that expires within an hour. Agents there cannot
git pushor callgh; the pipeline pushes and opens the PR itself. - Planning runs on a read-only clone with read-only tools. Repository code never executes in the control plane.
- Approval requires a chat user linked to a GitHub account that has write access to the repository.
- Teams requests are verified against the Bot Framework signing keys, the app id and the service URL. Slack uses Socket Mode, so it exposes no inbound endpoint.
- Prompt injection: repository content and issue text can contain instructions aimed at the agents. The runner's sandbox, the scoped token and the human review of every PR are the controls. Treat a Pullwright PR like any external contribution and review it before merging.